Refine
Year of publication
Document Type
- Article (61)
- conference proceeding (article) (52)
- conference proceeding (presentation, abstract) (33)
- Part of Periodical (5)
- Preprint (4)
- Book (1)
- Part of a Book (1)
Is part of the Bibliography
- no (157)
Keywords
- Bildgebendes Verfahren (18)
- Deep Learning (16)
- Diagnose (14)
- Künstliche Intelligenz (13)
- Artificial Intelligence (12)
- Maschinelles Lernen (12)
- Gehirn (11)
- Kernspintomografie (9)
- Dreidimensionale Bildverarbeitung (8)
- Registrierung <Bildverarbeitung> (8)
Institute
- Fakultät Informatik und Mathematik (152)
- Regensburg Medical Image Computing (ReMIC) (151)
- Regensburg Center of Biomedical Engineering - RCBE (38)
- Regensburg Center of Health Sciences and Technology - RCHST (28)
- Hochschulleitung/Hochschulverwaltung (5)
- Zentrum für Forschung und Transfer (ZFT ab 2024; vorher: IAFW) (5)
- Fakultät Sozial- und Gesundheitswissenschaften (2)
- Labor Empirische Sozialforschung (2)
- Labor für Technikfolgenabschätzung und Angewandte Ethik (LaTe) (2)
- Fakultät Maschinenbau (1)
Begutachtungsstatus
- peer-reviewed (78)
- begutachtet (1)
Purpose
MR-based attenuation correction (AC) will become an integral part of combined PET/MR systems. Here, we propose a toolbox to validate MR-AC of clinical PET/MRI data sets.
Methods
Torso scans of ten patients were acquired on a combined PET/CT and on a 1.5-T MRI system. MR-based attenuation data were derived from the CT following MR–CT image co-registration and subsequent histogram matching. PET images were reconstructed after CT- (PET/CT) and MR-based AC (PET/MRI). Lesion-to-background (L/B) ratios were estimated on PET/CT and PET/MRI.
Results
MR–CT histogram matching leads to a mean voxel intensity difference in the CT- and MR-based attenuation images of 12% (max). Mean differences between PET/MRI and PET/CT were 19% (max). L/B ratios were similar except for the lung where local misregistration and intensity transformation leads to a biased PET/MRI.
Conclusion
Our toolbox can be used to study pitfalls in MR-AC. We found that co-registration accuracy and pixel value transformation determine the accuracy of PET/MRI.
Mass spectrometric imaging (MSI) is a young innovative analytical technique and combines different fields of advanced mass spectrometry and biomedical research with the aim to provide maps of elements and molecules, complexes or fragments. Especially essential metals such as zinc, copper, iron and manganese play a functional role in signaling, metabolism and homeostasis of the cell. Due to the high degree of spatial organization of metals in biological systems their distribution analysis is of key interest in life sciences. We have developed analytical techniques termed BrainMet using laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS) imaging to measure the distribution of trace metals in biological tissues for biomedical research and feasibility studies—including bioaccumulation and bioavailability studies, ecological risk assessment and toxicity studies in humans and other organisms. The analytical BrainMet techniques provide quantitative images of metal distributions in brain tissue slices which can be combined with other imaging modalities such as photomicrography of native or processed tissue (histochemistry, immunostaining) and autoradiography or with in vivo techniques such as positron emission tomography or magnetic resonance tomography.
Prospective and instrumental developments will be discussed concerning the development of the metalloprotein microscopy using a laser microdissection (LMD) apparatus for specific sample introduction into an inductively coupled plasma mass spectrometer (LMD-ICP-MS) or an application of the near field effect in LA-ICP-MS (NF-LA-ICP-MS). These nano-scale mass spectrometric techniques provide improved spatial resolution down to the single cell level.
Local gray level dependencies of natural images can be modelled by means of co-occurrence matrices containing joint probabilities of gray-level pairs. Texture, however, is a resolution-dependent phenomenon and hence, classification depends on the chosen scale. Since there is no optimal scale for all textures we employ a multiscale approach that acquires textural features at several scales. Thus linear and nonlinear scale-spaces are analyzed by multiscale co-occurrence matrices that describe the statistical behavior of a texture in scale-space. Classification is then performed on the basis of texture features taken from the individual scale with the highest discriminatory power. By considering cross-scale occurrences of gray level pairs, the impact of filters on the feature is described and used for classification of natural textures. This novel method was found to improve classification rates of the common co-occurrence matrix approach on standard textures significantly.
Background:
Reliable, time- and cost-effective, and clinician-friendly diagnostic tools are cornerstones in facial palsy (FP) patient management. Different automated FP grading systems have been developed but revealed persisting downsides such as insufficient accuracy and cost-intensive hardware. We aimed to overcome these barriers and programmed an automated grading system for FP patients utilizing the House and Brackmann scale (HBS).
Methods:
Image datasets of 86 patients seen at the Department of Plastic, Hand, and Reconstructive Surgery at the University Hospital Regensburg, Germany, between June 2017 and May 2021, were used to train the neural network and evaluate its accuracy. Nine facial poses per patient were analyzed by the algorithm.
Results:
The algorithm showed an accuracy of 100%. Oversampling did not result in altered outcomes, while the direct form displayed superior accuracy levels when compared to the modular classification form (n = 86; 100% vs. 99%). The Early Fusion technique was linked to improved accuracy outcomes in comparison to the Late Fusion and sequential method (n = 86; 100% vs. 96% vs. 97%).
Conclusions:
Our automated FP grading system combines high-level accuracy with cost- and time-effectiveness. Our algorithm may accelerate the grading process in FP patients and facilitate the FP surgeon’s workflow.
In the field of computer- and robot-assisted minimally invasive surgery, enormous progress has been made in recent years based on the recognition of surgical instruments in endoscopic images. Especially the determination of the position and type of the instruments is of great interest here. Current work involves both spatial and temporal information with the idea, that the prediction of movement of surgical tools over time may improve the quality of final segmentations. The provision of publicly available datasets has recently encouraged the development of new methods, mainly based on deep learning. In this review, we identify datasets used for method development and evaluation, as well as quantify their frequency of use in the literature. We further present an overview of the current state of research regarding the segmentation and tracking of minimally invasive surgical instruments in endoscopic images. The paper focuses on methods that work purely visually without attached markers of any kind on the instruments, taking into account both single-frame segmentation approaches as well as those involving temporal information. A discussion of the reviewed literature is provided, highlighting existing shortcomings and emphasizing available potential for future developments. The publications considered were identified through the platforms Google Scholar, Web of Science, and PubMed. The search terms used were "instrument segmentation", "instrument tracking", "surgical tool segmentation", and "surgical tool tracking" and result in 408 articles published between 2015 and 2022 from which 109 were included using systematic selection criteria.
We investigate contrastive learning in a multi-task learning setting classifying and segmenting early Barrett’s cancer. How can contrastive learning be applied in a domain with few classes and low inter-class and inter-sample variance, potentially enabling image retrieval or image attribution? We introduce a data sampling strategy that mines per-lesion data for positive samples and keeps a queue of the recent projections as negative samples. We propose a masking strategy for the NT-Xent loss that keeps the negative set pure and removes samples from the same lesion. We show cohesion and uniqueness improvements of the proposed method in feature space. The introduction of the auxiliary objective does not affect the performance but adds the ability to indicate similarity between lesions. Therefore, the approach could enable downstream auto-documentation tasks on homogeneous medical image data.
BACKGROUND
Differentiation of high-flow from low-flow vascular malformations (VMs) is crucial for therapeutic management of this orphan disease.
OBJECTIVE
A convolutional neural network (CNN) was evaluated for differentiation of peripheral vascular malformations (VMs) on T2-weighted short tau inversion recovery (STIR) MRI.
METHODS
527 MRIs (386 low-flow and 141 high-flow VMs) were randomly divided into training, validation and test set for this single-center study. 1) Results of the CNN's diagnostic performance were compared with that of two expert and four junior radiologists. 2) The influence of CNN's prediction on the radiologists' performance and diagnostic certainty was evaluated. 3) Junior radiologists' performance after self-training was compared with that of the CNN.
RESULTS
Compared with the expert radiologists the CNN achieved similar accuracy (92% vs. 97%, p = 0.11), sensitivity (80% vs. 93%, p = 0.16) and specificity (97% vs. 100%, p = 0.50). In comparison to the junior radiologists, the CNN had a higher specificity and accuracy (97% vs. 80%, p < 0.001; 92% vs. 77%, p < 0.001). CNN assistance had no significant influence on their diagnostic performance and certainty. After self-training, the junior radiologists' specificity and accuracy improved and were comparable to that of the CNN.
CONCLUSIONS
Diagnostic performance of the CNN for differentiating high-flow from low-flow VM was comparable to that of expert radiologists. CNN did not significantly improve the simulated daily practice of junior radiologists, self-training was more effective.
Biological applications like vesicle membrane analysis involve the precise segmentation of 3D structures in noisy volumetric data, obtained by techniques like magnetic resonance imaging (MRI) or laser scanning microscopy (LSM). Dealing with such data is a challenging task and requires robust and accurate segmentation methods. In this article, we propose a novel energy model for 3D segmentation fusing various cues like regional intensity subdivision, edge alignment and orientation information. The uniqueness of the approach consists in the definition of a new anisotropic regularizer, which accounts for the unbalanced slicing of the measured volume data, and the generalization of an efficient numerical scheme for solving the arising minimization problem, based on linearization and fixed-point iteration. We show how the proposed energy model can be optimized globally by making use of recent continuous convex relaxation techniques. The accuracy and robustness of the presented approach are demonstrated by evaluating it on multiple real data sets and comparing it to alternative segmentation methods based on level sets. Although the proposed model is designed with focus on the particular application at hand, it is general enough to be applied to a variety of different segmentation tasks.
The field of AI is characterized by robust promises, astonishing successes, and remarkable breakthroughs. AI will play a major role in all domains of clinical medicine, but the role of AI in relation to the physician is not yet completely determined. The term artificial intelligence or AI is broad, and several different terms are used in this context that must be organized and demystified. This chapter will review the key concepts and methods of AI, and will introduce some of the different roles for AI in relation to the physician.
In the field of computer- and robot-assisted minimally invasive surgery, enormous progress has been made in recent years based on the recognition of surgical instruments in endoscopic images and videos. In particular, the determination of the position and type of instruments is of great interest. Current work involves both spatial and temporal information, with the idea that predicting the movement of surgical tools over time may improve the quality of the final segmentations. The provision of publicly available datasets has recently encouraged the development of new methods, mainly based on deep learning. In this review, we identify and characterize datasets used for method development and evaluation and quantify their frequency of use in the literature. We further present an overview of the current state of research regarding the segmentation and tracking of minimally invasive surgical instruments in endoscopic images and videos. The paper focuses on methods that work purely visually, without markers of any kind attached to the instruments, considering both single-frame semantic and instance segmentation approaches, as well as those that incorporate temporal information. The publications analyzed were identified through the platforms Google Scholar, Web of Science, and PubMed. The search terms used were “instrument segmentation”, “instrument tracking”, “surgical tool segmentation”, and “surgical tool tracking”, resulting in a total of 741 articles published between 01/2015 and 07/2023, of which 123 were included using systematic selection criteria. A discussion of the reviewed literature is provided, highlighting existing shortcomings and emphasizing the available potential for future developments.
Digital nerve lesions result in a loss of tactile sensation reflected by an anesthetic area (AA) at the radial or ulnar aspect of the respective digit. Yet, available tools to monitor the recovery of tactile sense have been criticized for their lack of validity. However, the precise quantification of AA dynamics by three-dimensional (3-D) imaging could serve as an accurate surrogate to monitor recovery following digital nerve repair.
For validation, AAs were marked on digits of healthy volunteers to simulate the AA of an impaired cutaneous innervation. Three dimensional models were composed from raw images that had been acquired with a 3-D camera (Vectra H2) to precisely quantify relative AA for each digit (3-D models, n= 80). Operator properties varied regarding individual experience in 3-D imaging and image processing. Additionally, the concept was applied in a clinical case study.
Images taken by experienced photographers were rated better quality (p< 0.001) and needed less processing time (p= 0.020). Quantification of the relative AA was neither altered significantly by experience levels of the photographer (p= 0.425) nor the image assembler (p= 0.749).
The proposed concept allows precise and reliable surface quantification of digits and can be performed consistently without relevant distortion by lack of examiner experience. Routine 3-D imaging of the AA has the great potential to provide visual evidence of various returning states of sensation and to convert sensory nerve recovery into a metric variable with high responsiveness to temporal progress.
We present the Regensburg Breast Shape Model (RBSM)—a 3D statistical shape model of the female breast built from 110 breast scans acquired in a standing position, and the first publicly available. Together with the model, a fully automated, pairwise surface registration pipeline used to establish dense correspondence among 3D breast scans is introduced. Our method is computationally efficient and requires only four landmarks to guide the registration process. A major challenge when modeling female breasts from surface-only 3D breast scans is the non-separability of breast and thorax. In order to weaken the strong coupling between breast and surrounding areas, we propose to minimize the variance outside the breast region as much as possible. To achieve this goal, a novel concept called breast probability masks (BPMs) is introduced. A BPM assigns probabilities to each point of a 3D breast scan, telling how likely it is that a particular point belongs to the breast area. During registration, we use BPMs to align the template to the target as accurately as possible inside the breast region and only roughly outside. This simple yet effective strategy significantly reduces the unwanted variance outside the breast region, leading to better statistical shape models in which breast shapes are quite well decoupled from the thorax. The RBSM is thus able to produce a variety of different breast shapes as independently as possible from the shape of the thorax. Our systematic experimental evaluation reveals a generalization ability of 0.17 mm and a specificity of 2.8 mm. To underline the expressiveness of the proposed model, we finally demonstrate in two showcase applications how the RBSM can be used for surgical outcome simulation and the prediction of a missing breast from the remaining one. Our model is available at https://www.rbsm.re-mic.de/.
Object
The consistency of parametric registration in multi-temporal magnetic resonance (MR) imaging studies was evaluated.
Materials and methods
Serial MRI scans of adult patients with a brain tumor (glioma) were aligned by parametric registration. The performance of low-order spatial alignment (6/9/12 degrees of freedom) of different 3D serial MR-weighted images is evaluated. A registration protocol for the alignment of all images to one reference coordinate system at baseline is presented. Registration results were evaluated for both, multimodal intra-timepoint and mono-modal multi-temporal registration. The latter case might present a challenge to automatic intensity-based registration algorithms due to ill-defined correspondences. The performance of our algorithm was assessed by testing the inverse registration consistency. Four different similarity measures were evaluated to assess consistency.
Results
Careful visual inspection suggests that images are well aligned, but their consistency may be imperfect. Sub-voxel inconsistency within the brain was found for allsimilarity measures used for parametric multi-temporal registration. T1-weighted images were most reliable for establishing spatial correspondence between different timepoints.
Conclusions
The parametric registration algorithm is feasible for use in this application. The sub-voxel resolution mean displacement error of registration transformations demonstrates that the algorithm converges to an almost identical solution for forward and reverse registration.
Chest X-Ray (CXR) images as part of a non-invasive diagnosis method are commonly used in today’s medical workflow. In traditional methods, physicians usually use their experience to interpret CXR images, however, there is a large interobserver variance. Computer vision may be used as a standard for assisted diagnosis. In this study, we applied an encoder-decoder neural network architecture for automatic lung region detection. We compared a three-class approach (left lung, right lung, background) and a two-class approach (lung, background). The differentiation of left and right lungs as direct result of a semantic segmentation on basis of neural nets rather than post-processing a lung-background segmentation is done here for the first time. Our evaluation was done on the NIH Chest X-ray dataset, from which 1736 images were extracted and manually annotated. We achieved 94:9% mIoU and 92% mIoU as segmentation quality measures for the two-class-model and the three-class-model, respectively. This result is very promising for the segmentation of lung regions having the simultaneous classification of left and right lung in mind.
3D rat brain tumors
(2005)
Clinical setting
Third space procedures such as endoscopic submucosal dissection (ESD) and peroral endoscopic myotomy (POEM) are complex minimally invasive techniques with an elevated risk for operator-dependent adverse events such as bleeding and perforation. This risk arises from accidental dissection into the muscle layer or through submucosal blood vessels as the submucosal cutting plane within the expanding resection site is not always apparent. Deep learning algorithms have shown considerable potential for the detection and characterization of gastrointestinal lesions. So-called AI – clinical decision support solutions (AI-CDSS) are commercially available for polyp detection during colonoscopy. Until now, these computer programs have concentrated on diagnostics whereas an AI-CDSS for interventional endoscopy has not yet been introduced. We aimed to develop an AI-CDSS („Smart ESD“) for real-time intra-procedural detection and delineation of blood vessels, tissue structures and endoscopic instruments during third-space endoscopic procedures.
Characteristics of Smart ESD
An AI-CDSS was invented that delineates blood vessels, tissue structures and endoscopic instruments during third-space endoscopy in real-time. The output can be displayed by an overlay over the endoscopic image with different modes of visualization, such as a color-coded semitransparent area overlay, or border tracing (demonstration video). Hereby the optimal layer for dissection can be visualized, which is close above or directly at the muscle layer, depending on the applied technique (ESD or POEM). Furthermore, relevant blood vessels (thickness> 1mm) are delineated. Spatial proximity between the electrosurgical knife and a blood vessel triggers a warning signal. By this guidance system, inadvertent dissection through blood vessels could be averted.
Technical specifications
A DeepLabv3+ neural network architecture with KSAC and a 101-layer ResNeSt backbone was used for the development of Smart ESD. It was trained and validated with 2565 annotated still images from 27 full length third-space endoscopic videos. The annotation classes were blood vessel, submucosal layer, muscle layer, electrosurgical knife and endoscopic instrument shaft. A test on a separate data set yielded an intersection over union (IoU) of 68%, a Dice Score of 80% and a pixel accuracy of 87%, demonstrating a high overlap between expert and AI segmentation. Further experiments on standardized video clips showed a mean vessel detection rate (VDR) of 85% with values of 92%, 70% and 95% for POEM, rectal ESD and esophageal ESD respectively. False positive measurements occurred 0.75 times per minute. 7 out of 9 vessels which caused intraprocedural bleeding were caught by the algorithm, as well as both vessels which required hemostasis via hemostatic forceps.
Future perspectives
Smart ESD performed well for vessel and tissue detection and delineation on still images, as well as on video clips. During a live demonstration in the endoscopy suite, clinical applicability of the innovation was examined. The lag time for processing of the live endoscopic image was too short to be visually detectable for the interventionist. Even though the algorithm could not be applied during actual dissection by the interventionist, Smart ESD appeared readily deployable during visual assessment by ESD experts. Therefore, we plan to conduct a clinical trial in order to obtain CE-certification of the algorithm. This new technology may improve procedural safety and speed, as well as training of modern minimally invasive endoscopic resection techniques.
Aims
Barrett´s esophagus related neoplasia (BERN) is difficult to detect and characterize during endoscopy, even for expert endoscopists. We aimed to assess the add-on effect of an Artificial Intelligence (AI) algorithm (Barrett-Ampel) as a decision support system (DSS) for non-expert endoscopists in the evaluation of Barrett’s esophagus (BE) and BERN.
Methods
Twelve videos with multimodal imaging white light (WL), narrow-band imaging (NBI), texture and color enhanced imaging (TXI) of histologically confirmed BE and BERN were assessed by expert and non-expert endoscopists. For each video, endoscopists were asked to identify the area of BERN and decide on the biopsy spot. Videos were assessed by the AI algorithm and regions of BERN were highlighted in real-time by a transparent overlay. Finally, endoscopists were shown the AI videos and asked to either confirm or change their initial decision based on the AI support.
Results
Barrett-Ampel correctly identified all areas of BERN, irrespective of the imaging modality (WL, NBI, TXI), but misinterpreted two inflammatory lesions (Accuracy=75%). Expert endoscopists had a similar performance (Accuracy=70,8%), while non-experts had an accuracy of 58.3%. When AI was implemented as a DSS, non-expert endoscopists improved their diagnostic accuracy to 75%.
Conclusions
AI may have the potential to support non-expert endoscopists in the assessment of videos of BE and BERN. Limitations of this study include the low number of videos used. Randomized clinical trials in a real-life setting should be performed to confirm these results.
ARTIFICIAL INTELLIGENCE (AI) – ASSISTED VESSEL AND TISSUE RECOGNITION IN THIRD-SPACE ENDOSCOPY
(2022)
Aims
Third-space endoscopy procedures such as endoscopic submucosal dissection (ESD) and peroral endoscopic myotomy (POEM) are complex interventions with elevated risk of operator-dependent adverse events, such as intra-procedural bleeding and perforation. We aimed to design an artificial intelligence clinical decision support solution (AI-CDSS, “Smart ESD”) for the detection and delineation of vessels, tissue structures, and instruments during third-space endoscopy procedures.
Methods
Twelve full-length third-space endoscopy videos were extracted from the Augsburg University Hospital database. 1686 frames were annotated for the following categories: Submucosal layer, blood vessels, electrosurgical knife and endoscopic instrument. A DeepLabv3+neural network with a 101-layer ResNet backbone was trained and validated internally. Finally, the ability of the AI system to detect visible vessels during ESD and POEM was determined on 24 separate video clips of 7 to 46 seconds duration and showing 33 predefined vessels. These video clips were also assessed by an expert in third-space endoscopy.
Results
Smart ESD showed a vessel detection rate (VDR) of 93.94%, while an average of 1.87 false positive signals were recorded per minute. VDR of the expert endoscopist was 90.1% with no false positive findings. On the internal validation data set using still images, the AI system demonstrated an Intersection over Union (IoU), mean Dice score and pixel accuracy of 63.47%, 76.18% and 86.61%, respectively.
Conclusions
This is the first AI-CDSS aiming to mitigate operator-dependent limitations during third-space endoscopy. Further clinical trials are underway to better understand the role of AI in such procedures.
Utility of Smartphone-based Three-dimensional Surface Imaging for Digital Facial Anthropometry
(2024)
Background
The utilization of three-dimensional (3D) surface imaging for facial anthropometry is a significant asset for patients undergoing maxillofacial surgery. Notably, there have been recent advancements in smartphone technology that enable 3D surface imaging.
In this study, anthropometric assessments of the face were performed using a smartphone and a sophisticated 3D surface imaging system.
Methods
30 healthy volunteers (15 females and 15 males) were included in the study. An iPhone 14 Pro (Apple Inc., USA) using the application 3D Scanner App (Laan Consulting Corp., USA) and the Vectra M5 (Canfield Scientific, USA) were employed to create 3D surface models. For each participant, 19 anthropometric measurements were conducted on the 3D surface models. Subsequently, the anthropometric measurements generated by the two approaches were compared. The statistical techniques employed included the paired t-test, paired Wilcoxon signed-rank test, Bland–Altman analysis, and calculation of the intraclass correlation coefficient (ICC).
Results
All measurements showed excellent agreement between smartphone-based and Vectra M5-based measurements (ICC between 0.85 and 0.97). Statistical analysis revealed no statistically significant differences in the central tendencies for 17 of the 19 linear measurements. Despite the excellent agreement found, Bland–Altman analysis revealed that the 95% limits of agreement between the two methods exceeded ±3 mm for the majority of measurements.
Conclusion
Digital facial anthropometry using smartphones can serve as a valuable supplementary tool for surgeons, enhancing their communication with patients. However, the proposed data suggest that digital facial anthropometry using smartphones may not yet be suitable for certain diagnostic purposes that require high accuracy.
Aims
Celiac disease (CD) is a complex condition caused by an autoimmune reaction to ingested gluten. Due to its polymorphic manifestation and subtle endoscopic presentation, the diagnosis is difficult and thus the disorder is underreported. We aimed to use deep learning to identify celiac disease on endoscopic images of the small bowel.
Methods
Patients with small intestinal histology compatible with CD (MARSH classification I-III) were extracted retrospectively from the database of Augsburg University hospital. They were compared to patients with no clinical signs of CD and histologically normal small intestinal mucosa. In a first step MARSH III and normal small intestinal mucosa were differentiated with the help of a deep learning algorithm. For this, the endoscopic white light images were divided into five equal-sized subsets. We avoided splitting the images of one patient into several subsets. A ResNet-50 model was trained with the images from four subsets and then validated with the remaining subset. This process was repeated for each subset, such that each subset was validated once. Sensitivity, specificity, and harmonic mean (F1) of the algorithm were determined.
Results
The algorithm showed values of 0.83, 0.88, and 0.84 for sensitivity, specificity, and F1, respectively. Further data showing a comparison between the detection rate of the AI model and that of experienced endoscopists will be available at the time of the upcoming conference.
Conclusions
We present the first clinical report on the use of a deep learning algorithm for the detection of celiac disease using endoscopic images. Further evaluation on an external data set, as well as in the detection of CD in real-time, will follow. However, this work at least suggests that AI can assist endoscopists in the endoscopic diagnosis of CD, and ultimately may be able to do a true optical biopsy in live-time.
The early diagnosis of cancer in Barrett’s esophagus is crucial for improving the prognosis. However, identifying Barrett’s esophagus-related neoplasia (BERN) is challenging, even for experts [1]. Four-quadrant biopsies may improve the detection of neoplasia, but they can be associated with sampling errors. The application of artificial intelligence (AI) to the assessment of Barrett’s esophagus could improve the diagnosis of BERN, and this has been demonstrated in both preclinical and clinical studies [2] [3].
In this video demonstration, we show the accurate detection and delineation of BERN in two patients ([Video 1]). In part 1, the AI system detects a mucosal cancer about 20 mm in size and accurately delineates the lesion in both white-light and narrow-band imaging. In part 2, a small island of BERN with high-grade dysplasia is detected and delineated in white-light, narrow-band, and texture and color enhancement imaging. The video shows the results using a transparent overlay of the mucosal cancer in real time as well as a full segmentation preview. Additionally, the optical flow allows for the assessment of endoscope movement, something which is inversely related to the reliability of the AI prediction. We demonstrate that multimodal imaging can be applied to the AI-assisted detection and segmentation of even small focal lesions in real time.
Real-time computational speed and a high degree of precision are requirements for computer-assisted interventions. Applying a segmentation network to a medical video processing task can introduce significant inter-frame prediction noise. Existing approaches can reduce inconsistencies by including temporal information but often impose requirements on the architecture or dataset. This paper proposes a method to include temporal information in any segmentation model and, thus, a technique to improve video segmentation performance without alterations during training or additional labeling. With Motion-Corrected Moving Average, we refine the exponential moving average between the current and previous predictions. Using optical flow to estimate the movement between consecutive frames, we can shift the prior term in the moving-average calculation to align with the geometry of the current frame. The optical flow calculation does not require the output of the model and can therefore be performed in parallel, leading to no significant runtime penalty for our approach. We evaluate our approach on two publicly available segmentation datasets and two proprietary endoscopic datasets and show improvements over a baseline approach.
We propose an automatic approach for early detection of adenocarcinoma in the esophagus. High-definition endoscopic images (50 cancer, 50 Barrett) are partitioned into a dataset containing approximately equal amounts of patches showing cancerous and non-cancerous regions. A deep convolutional neural network is adapted to the data using a transfer learning approach. The final classification of an image is determined by at least one patch, for which the probability being a cancer patch exceeds a given threshold. The model was evaluated with leave one patient out cross-validation. With sensitivity and specificity of 0.94 and 0.88, respectively, our findings improve recently published results on the same image data base considerably. Furthermore, the visualization of the class probabilities of each individual patch indicates, that our approach might be extensible to the segmentation domain.
Celiac disease is an autoimmune disorder caused by gluten that results in an inflammatory response of the small intestine.We investigated whether celiac disease can be detected using endoscopic images through a deep learning approach. The results show that additional clinical parameters can improve the classification accuracy. In this work, we distinguished between healthy tissue and Marsh III, according to the Marsh score system. We first trained a baseline network to classify endoscopic images of the small bowel into these two classes and then augmented the approach with a multimodality component that took the antibody status into account.
Einleitung
Die sichere Detektion und Charakterisierung von Barrett-Ösophagus assoziierten Neoplasien (BERN) stellt selbst für erfahrene Endoskopiker eine Herausforderung dar.
Ziel
Ziel dieser Studie ist es, den Add-on Effekt eines künstlichen Intelligenz (KI) Systems (Barrett-Ampel) als Entscheidungsunterstüzungssystem für Endoskopiker ohne Expertise bei der Untersuchung von BERN zu evaluieren.
Material und Methodik
Zwölf Videos in „Weißlicht“ (WL), „narrow-band imaging“ (NBI) und „texture and color enhanced imaging“ (TXI) von histologisch bestätigten Barrett-Metaplasien oder BERN wurden von Experten und Untersuchern ohne Barrett-Expertise evaluiert. Die Probanden wurden dazu aufgefordert in den Videos auftauchende BERN zu identifizieren und gegebenenfalls die optimale Biopsiestelle zu markieren. Unser KI-System wurde demselben Test unterzogen, wobei dieses BERN in Echtzeit segmentierte und farblich von umliegendem Epithel differenzierte. Anschließend wurden den Probanden die Videos mit zusätzlicher KI-Unterstützung gezeigt. Basierend auf dieser neuen Information, wurden die Probanden zu einer Reevaluation ihrer initialen Beurteilung aufgefordert.
Ergebnisse
Die „Barrett-Ampel“ identifizierte unabhängig von den verwendeten Darstellungsmodi (WL, NBI, TXI) alle BERN. Zwei entzündlich veränderte Läsionen wurden fehlinterpretiert (Genauigkeit=75%). Während Experten vergleichbare Ergebnisse erzielten (Genauigkeit=70,8%), hatten Endoskopiker ohne Expertise bei der Beurteilung von Barrett-Metaplasien eine Genauigkeit von lediglich 58,3%. Wurden die nicht-Experten allerdings von unserem KI-System unterstützt, erreichten diese eine Genauigkeit von 75%.
Zusammenfassung
Unser KI-System hat das Potential als Entscheidungsunterstützungssystem bei der Differenzierung zwischen Barrett-Metaplasie und BERN zu fungieren und so Endoskopiker ohne entsprechende Expertise zu assistieren. Eine Limitation dieser Studie ist die niedrige Anzahl an eingeschlossenen Videos. Um die Ergebnisse dieser Studie zu bestätigen, müssen randomisierte kontrollierte klinische Studien durchgeführt werden.
Einleitung
Third-Space Interventionen wie die endoskopische Submukosadissektion (ESD) und die perorale endoskopische Myotomie (POEM) sind technisch anspruchsvoll und mit einem erhöhten Risiko für intraprozedurale Komplikationen wie Blutung oder Perforation assoziiert. Moderne Computerprogramme zur Unterstützung bei diagnostischen Entscheidungen werden unter Einsatz von künstlicher Intelligenz (KI) in der Endoskopie bereits erfolgreich eingesetzt. Ziel der vorliegenden Arbeit war es, relevante anatomische Strukturen mithilfe eines Deep-Learning Algorithmus zu detektieren und segmentieren, um die Sicherheit und Anwendbarkeit von ESD und POEM zu erhöhen.
Methoden
Zwölf Videoaufnahmen in voller Länge von Third-Space Endoskopien wurden aus der Datenbank des Universitätsklinikums Augsburg extrahiert. 1686 Einzelbilder wurden für die Kategorien Submukosa, Blutgefäß, Dissektionsmesser und endoskopisches Instrument annotiert und segmentiert. Mit diesem Datensatz wurde ein DeepLabv3+neuronales Netzwerk auf der Basis eines ResNet mit 101 Schichten trainiert und intern anhand der Parameter Intersection over Union (IoU), Dice Score und Pixel Accuracy validiert. Die Fähigkeit des Algorithmus zur Gefäßdetektion wurde anhand von 24 Videoclips mit einer Spieldauer von 7 bis 46 Sekunden mit 33 vordefinierten Gefäßen evaluiert. Anhand dieses Tests wurde auch die Gefäßdetektionsrate eines Experten in der Third-Space Endoskopie ermittelt.
Ergebnisse
Der Algorithmus zeigte eine Gefäßdetektionsrate von 93,94% mit einer mittleren Rate an falsch positiven Signalen von 1,87 pro Minute. Die Gefäßdetektionsrate des Experten lag bei 90,1% ohne falsch positive Ergebnisse. In der internen Validierung an Einzelbildern wurde eine IoU von 63,47%, ein mittlerer Dice Score von 76,18% und eine Pixel Accuracy von 86,61% ermittelt.
Zusammenfassung
Dies ist der erste KI-Algorithmus, der für den Einsatz in der therapeutischen Endoskopie entwickelt wurde. Präliminäre Ergebnisse deuten auf eine mit Experten vergleichbare Detektion von Gefäßen während der Untersuchung hin. Weitere Untersuchungen sind nötig, um die Leistung des Algorithmus im Vergleich zum Experten genauer zu eruieren sowie einen möglichen klinischen Nutzen zu ermitteln.
Introduction
We present a clinical case showing the real-time detection, characterization and delineation of an early Barrett’s cancer using AI.
Patients and methods
A 70-year old patient with a long-segment Barrett’s esophagus (C5M7) was assessed with an AI algorithm.
Results
The AI system detected a 10 mm focal lesion and AI characterization predicted cancer with a probability of >90%. After ESD resection, histopathology showed mucosal adenocarcinoma (T1a (m), R0) confirming AI diagnosis.
Conclusion
We demonstrate the real-time AI detection, characterization and delineation of a small and early mucosal Barrett’s cancer.
In 2015 we began a sub-challenge at the EndoVis workshop at MICCAI in Munich using endoscope images of exvivo tissue with automatically generated annotations from robot forward kinematics and instrument CAD models. However, the limited background variation and simple motion rendered the dataset uninformative in learning about which techniques would be suitable for segmentation in real surgery. In 2017, at the same workshop in Quebec we introduced the robotic instrument segmentation dataset with 10 teams participating in the challenge to perform binary, articulating parts and type segmentation of da Vinci instruments. This challenge included realistic instrument motion and more complex porcine tissue as background and was widely addressed with modfications on U-Nets and other popular CNN architectures [1].
In 2018 we added to the complexity by introducing a set of anatomical objects and medical devices to the segmented classes. To avoid over-complicating the challenge, we continued with porcine data which is dramatically simpler than human tissue due to the lack of fatty tissue occluding many organs.
Einleitung
Third space Endoskopieprozeduren wie die endoskopische Submukosadissektion (ESD) und die perorale endoskopische Myotomie (POEM) sind technisch anspruchsvoll und gehen mit untersucherabhängigen Komplikationen wie Blutungen und Perforationen einher. Grund hierfür ist die unabsichtliche Durchschneidung von submukosalen Blutgefäßen ohne präemptive Koagulation.
Ziele
Die Forschungsfrage, ob ein KI-Algorithmus die intraprozedurale Gefäßerkennung bei ESD und POEM unterstützen und damit Komplikationen wie Blutungen verhindern könnte, erscheint in Anbetracht des erfolgreichen Einsatzes von KI bei der Erkennung von Kolonpolypen interessant.
Methoden
Auf 5470 Einzelbildern von 59 third space Endoscopievideos wurden submukosale Blutgefäße annotiert. Zusammen mit weiteren 179.681 nicht-annotierten Bildern wurde ein DeepLabv3+neuronales Netzwerk mit dem ECMT-Verfahren für semi-supervised learning trainiert, um Blutgefäße in Echtzeit erkennen zu können. Für die Evaluation wurde ein Videotest mit 101 Videoclips aus 15 vom Trainingsdatensatz separaten Prozeduren mit 200 vordefinierten Gefäßen erstellt. Die Gefäßdetektionsrate, -zeit und -dauer, definiert als der Prozentsatz an Einzelbildern eines Videos bezogen auf den Goldstandard, auf denen ein definiertes Gefäß erkannt wurde, wurden erhoben. Acht erfahrene Endoskopiker wurden mithilfe dieses Videotests im Hinblick auf Gefäßdetektion getestet, wobei eine Hälfte der Videos nativ, die andere Hälfte nach Markierung durch den KI-Algorithmus angesehen wurde.
Ergebnisse
Der mittlere Dice Score des Algorithmus für Blutgefäße war 68%. Die mittlere Gefäßdetektionsrate im Videotest lag bei 94% (96% für ESD; 74% für POEM). Die mediane Gefäßdetektionszeit des Algorithmus lag bei 0,32 Sekunden (0,3 Sekunden für ESD; 0,62 Sekunden für POEM). Die mittlere Gefäßdetektionsdauer lag bei 59,1% (60,6% für ESD; 44,8% für POEM) des Goldstandards. Alle Endoskopiker hatten mit KI-Unterstützung eine höhere Gefäßdetektionsrate als ohne KI. Die mittlere Gefäßdetektionsrate ohne KI lag bei 56,4%, mit KI bei 71,2% (p<0.001).
Schlussfolgerung
KI-Unterstützung war mit einer statistisch signifikant höheren Gefäßdetektionsrate vergesellschaftet. Die mediane Gefäßdetektionszeit von deutlich unter einer Sekunde sowie eine Gefäßdetektionsdauer von größer 50% des Goldstandards wurden für den klinischen Einsatz als ausreichend erachtet. In prospektiven Anwendungsstudien sollte der KI-Algorithmus auf klinische Relevanz getestet werden.
Barrett-Ampel
(2022)
Hintergrund
Adenokarzinome des Ösophagus sind bis heute mit einer infausten Prognose vergesellschaftet (1). Obwohl Endoskopiker mit Barrett-Ösophagus als Präkanzerose konfrontiert werden, ist vor allem für nicht-Experten die Differenzierung zwischen Barrett-Ösophagus ohne Dysplasie und assoziierten Neoplasien mitunter schwierig. Existierende Biopsieprotokolle (z.B. Seattle Protokoll) sind oftmals unzuverlässig (2). Eine frühzeitige Diagnose des Adenokarzinoms ist allerdings von fundamentaler Bedeutung für die Prognose des Patienten.
Forschungsansatz
Auf der Grundlage dieser Problematik, entwickelten wir in Kooperation mit dem Forschungslabor „Regensburg Medical Image Computing (ReMIC)“ der OTH Regensburg ein auf künstlicher Intelligenz (KI) basiertes Entscheidungsunterstützungssystem (CDSS). Das auf einer DeepLabv3+ neuronalen Netzwerkarchitektur basierende CDSS differenziert mittels Mustererkennung Barrett- Ösophagus ohne Dysplasie von Barrett-Ösophagus mit Dysplasie bzw. Neoplasie („Klassifizierung“). Hierbei werden gemittelte Ausgabewahrscheinlichkeiten mit einem vom Benutzer definierten Schwellenwert verglichen. Für Vorhersagen, die den Schwellenwert überschreiten, berechnen wir die Kontur der Region und die Fläche. Sobald die vorhergesagte Läsion eine bestimmte Größe in der Eingabe überschreitet, heben wir sie und ihren Umriss hervor. So ermöglicht eine farbkodierte Visualisierung eine Abgrenzung zwischen Dysplasie bzw. Neoplasie und normalem Barrett-Epithel („Segmentierung“).
In einer Studie an Bildern in „Weißlicht“ (WL) und „Narrow Band Imaging“ (NBI) demonstrierten wir eine Sensitivität von mehr als 90% und eine Spezifität von mehr als 80% (3). In einem nächsten Schritt, differenzierte unser KI-Algorithmus Barrett- Metaplasien von assoziierten Neoplasien anhand von zufällig abgegriffenen Bildern in Echtzeit mit einer Accuracy von 89.9% (4). Darauf folgend, entwickelten wir unser System dahingehend weiter, dass unser Algorithmus nun auch dazu in der Lage ist, Untersuchungsvideos in WL, NBI und „Texture and Color Enhancement Imaging“ (TXI) in Echtzeit zu analysieren (5).
Aktuell führen wir eine Studie in einem randomisiert-kontrollierten Ansatz an unveränderten Untersuchungsvideos in WL, NBI und TXI durch.
Ausblick
Um Patienten mit aus Barrett-Metaplasien resultierenden Neoplasien frühestmöglich an „High-Volume“-Zentren überweisen zu können, soll unser KI-Algorithmus zukünftig vor allem Endoskopiker ohne extensive Erfahrung bei der Beurteilung von Barrett- Ösophagus in der Krebsfrüherkennung unterstützen.
Einleitung
Übermäßige Bewegung im Bild kann die Performance von auf künstlicher Intelligenz (KI) basierenden klinischen Entscheidungsunterstützungssystemen (CDSS) reduzieren. Optical Flow (OF) ist eine Methode zur Lokalisierung und Quantifizierung von Bewegungen zwischen aufeinanderfolgenden Bildern.
Ziel
Ziel ist es, die Mensch-Computer-Interaktion (HCI) zu verbessern und Endoskopiker die unser KI-System „Barrett-Ampel“ zur Unterstützung bei der Beurteilung von Barrett-Ösophagus (BE) verwenden, ein Echtzeit-Feedback zur aktuellen Datenqualität anzubieten.
Methodik
Dazu wurden unveränderte Videos in „Weißlicht“ (WL), „Narrow Band Imaging“ (NBI) und „Texture and Color Enhancement Imaging“ (TXI) von acht endoskopischen Untersuchungen von histologisch gesichertem BE und mit Barrett-Ösophagus assoziierten Neoplasien (BERN) durch unseren KI-Algorithmus analysiert. Der zur Bewertung der Bildqualität verwendete OF beinhaltete die mittlere Magnitude und die Entropie des Histogramms der Winkel. Frames wurden automatisch extrahiert, wenn die vordefinierten Schwellenwerte von 3,0 für die mittlere Magnitude und 9,0 für die Entropie des Histogramms der Winkel überschritten wurden. Experten sahen sich zunächst die Videos ohne KI-Unterstützung an und bewerteten, ob Störfaktoren die Sicherheit mit der eine Diagnose im vorliegenden Fall gestellt werden kann negativ beeinflussen. Anschließend überprüften sie die extrahierten Frames.
Ergebnis
Gleichmäßige Bewegung in eine Richtung, wie etwa beim Vorschieben des Endoskops, spiegelte sich, bei insignifikant veränderter Entropie, in einer Erhöhung der Magnitude wider. Chaotische Bewegung, zum Beispiel während dem Spülen, war mit erhöhter Entropie assoziiert. Insgesamt war eine unruhige endoskopische Darstellung, Flüssigkeit sowie übermäßige Ösophagusmotilität mit erhöhtem OF assoziiert und korrelierte mit der Meinung der Experten über die Qualität der Videos. Der OF und die subjektive Wahrnehmung der Experten über die Verwertbarkeit der vorliegenden Bildsequenzen korrelierten direkt proportional. Wenn die vordefinierten Schwellenwerte des OF überschritten wurden, war die damit verbundene Bildqualität in 94% der Fälle für eine definitive Interpretation auch für Experten unzureichend.
Schlussfolgerung
OF hat das Potenzial Endoskopiker ein Echtzeit-Feedback über die Qualität des Dateninputs zu bieten und so nicht nur die HCI zu verbessern, sondern auch die optimale Performance von KI-Algorithmen zu ermöglichen.
Aims
AI has proven great potential in assisting endoscopists in diagnostics, however its role in therapeutic endoscopy remains unclear. Endoscopic submucosal dissection (ESD) is a technically demanding intervention with a slow learning curve and relevant risks like bleeding and perforation. Therefore, we aimed to develop an algorithm for the real-time detection and delineation of relevant structures during third-space endoscopy.
Methods
5470 still images from 59 full length videos (47 ESD, 12 POEM) were annotated. 179681 additional unlabeled images were added to the training dataset. Consequently, a DeepLabv3+ neural network architecture was trained with the ECMT semi-supervised algorithm (under review elsewhere). Evaluation of vessel detection was performed on a dataset of 101 standardized video clips from 15 separate third-space endoscopy videos with 200 predefined blood vessels.
Results
Internal validation yielded an overall mean Dice score of 85% (68% for blood vessels, 86% for submucosal layer, 88% for muscle layer). On the video test data, the overall vessel detection rate (VDR) was 94% (96% for ESD, 74% for POEM). The median overall vessel detection time (VDT) was 0.32 sec (0.3 sec for ESD, 0.62 sec for POEM).
Conclusions
Evaluation of the developed algorithm on a video test dataset showed high VDR and quick VDT, especially for ESD. Further research will focus on a possible clinical benefit of the AI application for VDR and VDT during third-space endoscopy.
Computer-aided diagnosis using deep learning in the evaluation of early oesophageal adenocarcinoma
(2019)
Computer-aided diagnosis using deep learning (CAD-DL) may be an instrument to improve endoscopic assessment of Barrett’s oesophagus
(BE) and early oesophageal adenocarcinoma (EAC). Based on still images from two databases, the diagnosis of EAC by CAD-DL reached sensitivities/specificities of 97%/88% (Augsburg data) and 92%/100% (Medical Image Computing and Computer-Assisted Intervention [MICCAI]
data) for white light (WL) images and 94%/80% for narrow band images (NBI) (Augsburg data), respectively. Tumour margins delineated by
experts into images were detected satisfactorily with a Dice coefficient (D) of 0.72. This could be a first step towards CAD-DL for BE assessment. If developed further, it could become a useful
adjunctive tool for patient management.
Barrett's esophagus denotes a disorder in the digestive system that affects the esophagus' mucosal cells, causing reflux, and showing potential convergence to esophageal adenocarcinoma if not treated in initial stages. Thus, fast and reliable computer-aided diagnosis becomes considerably welcome. Nevertheless, such approaches usually suffer from imbalanced datasets, which can be addressed through Generative Adversarial Networks (GANs). Such techniques generate realistic images based on observed samples, even though at the cost of a proper selection of its hyperparameters. Many works employed a class of nature-inspired algorithms called metaheuristics to tackle the problem considering distinct deep learning approaches. Therefore, this paper's main contribution is to introduce metaheuristic techniques to fine-tune GANs in the context of Barrett's esophagus identification, as well as to investigate the feasibility of generating high-quality synthetic images for early-cancer assisted identification.
Even though artificial intelligence and machine learning have demonstrated remarkable performances in medical image computing, their level of accountability and transparency must be provided in such evaluations. The reliability related to machine learning predictions must be explained and interpreted, especially if diagnosis support is addressed. For this task, the black-box nature of deep learning techniques must be lightened up to transfer its promising results into clinical practice. Hence, we aim to investigate the use of explainable artificial intelligence techniques to quantitatively highlight discriminative regions during the classification of earlycancerous tissues in Barrett’s esophagus-diagnosed patients. Four Convolutional Neural Network models (AlexNet, SqueezeNet, ResNet50, and VGG16) were analyzed using five different interpretation techniques (saliency, guided backpropagation, integrated gradients, input × gradients, and DeepLIFT) to compare their agreement with experts’ previous annotations of cancerous tissue. We could show that saliency attributes match best with the manual experts’ delineations. Moreover, there is moderate to high correlation between the sensitivity of a model and the human-and-computer agreement. The results also lightened that the higher the model’s sensitivity, the stronger the correlation of human and computational segmentation agreement. We observed a relevant relation between computational learning and experts’ insights, demonstrating how human knowledge may influence the correct computational learning.
Aims:
The delineation of outer margins of early Barrett's cancer can be challenging even for experienced endoscopists. Artificial intelligence (AI) could assist endoscopists faced with this task. As of date, there is very limited experience in this domain. In this study, we demonstrate the measure of overlap (Dice coefficient = D) between highly experienced Barrett endoscopists and an AI system in the delineation of cancer margins (segmentation task).
Methods:
An AI system with a deep convolutional neural network (CNN) was trained and tested on high-definition endoscopic images of early Barrett's cancer (n = 33) and normal Barrett's mucosa (n = 41). The reference standard for the segmentation task were the manual delineations of tumor margins by three highly experienced Barrett endoscopists. Training of the AI system included patch generation, patch augmentation and adjustment of the CNN weights. Then, the segmentation results from patch classification and thresholding of the class probabilities. Segmentation results were evaluated using the Dice coefficient (D).
Results:
The Dice coefficient (D) which can range between 0 (no overlap) and 1 (complete overlap) was computed only for images correctly classified by the AI-system as cancerous. At a threshold of t = 0.5, a mean value of D = 0.72 was computed.
Conclusions:
AI with CNN performed reasonably well in the segmentation of the tumor region in Barrett's cancer, at least when compared with expert Barrett's endoscopists. AI holds a lot of promise as a tool for better visualization of tumor margins but may need further improvement and enhancement especially in real-time settings.
The growing number of publications on the application of artificial intelligence (AI) in medicine underlines the enormous importance and potential of this emerging field of research.
In gastrointestinal endoscopy, AI has been applied to all segments of the gastrointestinal tract most importantly in the detection and characterization of colorectal polyps. However, AI research has been published also in the stomach and esophagus for both neoplastic and non-neoplastic disorders.
The various technical as well as medical aspects of AI, however, remain confusing especially for non-expert physicians.
This physician-engineer co-authored review explains the basic technical aspects of AI and provides a comprehensive overview of recent publications on AI in gastrointestinal endoscopy. Finally, a basic insight is offered into understanding publications on AI in gastrointestinal endoscopy.
Based on previous work by our group with manual annotation of visible Barrett oesophagus (BE) cancer images, a real-time deep learning artificial intelligence (AI) system was developed. While an expert endoscopist conducts the endoscopic assessment of BE, our AI system captures random images from the real-time camera livestream and provides a global prediction (classification), as well as a dense prediction (segmentation) differentiating accurately between normal BE and early oesophageal adenocarcinoma (EAC). The AI system showed an accuracy of 89.9% on 14 cases with neoplastic BE.
Semantic segmentation is an essential task in medical imaging research. Many powerful deep-learning-based approaches can be employed for this problem, but they are dependent on the availability of an expansive labeled dataset. In this work, we augment such supervised segmentation models to be suitable for learning from unlabeled data. Our semi-supervised approach, termed Error-Correcting Mean-Teacher, uses an exponential moving average model like the original Mean Teacher but introduces our new paradigm of error correction. The original segmentation network is augmented to handle this secondary correction task. Both tasks build upon the core feature extraction layers of the model. For the correction task, features detected in the input image are fused with features detected in the predicted segmentation and further processed with task-specific decoder layers. The combination of image and segmentation features allows the model to correct present mistakes in the given input pair. The correction task is trained jointly on the labeled data. On unlabeled data, the exponential moving average of the original network corrects the student’s prediction. The combined outputs of the students’ prediction with the teachers’ correction form the basis for the semi-supervised update. We evaluate our method with the 2017 and 2018 Robotic Scene Segmentation data, the ISIC 2017 and the BraTS 2020 Challenges, a proprietary Endoscopic Submucosal Dissection dataset, Cityscapes, and Pascal VOC 2012. Additionally, we analyze the impact of the individual components and examine the behavior when the amount of labeled data varies, with experiments performed on two distinct segmentation architectures. Our method shows improvements in terms of the mean Intersection over Union over the supervised baseline and competing methods. Code is available at https://github.com/CloneRob/ECMT.
Background and aims: The accurate differentiation between T1a and T1b Barrett’s cancer has both therapeutic and prognostic implications but is challenging even for experienced physicians. We trained an Artificial Intelligence (AI) system on the basis of deep artificial neural networks (deep learning) to differentiate between T1a and T1b Barrett’s cancer white-light images.
Methods: Endoscopic images from three tertiary care centres in Germany were collected retrospectively. A deep learning system was trained and tested using the principles of cross-validation. A total of 230 white-light endoscopic images (108 T1a and 122 T1b) was evaluated with the AI-system. For comparison, the images were also classified by experts specialized in endoscopic diagnosis and treatment of Barrett’s cancer.
Results: The sensitivity, specificity, F1 and accuracy of the AI-system in the differentiation between T1a and T1b cancer lesions was 0.77, 0.64, 0.73 and 0.71, respectively. There was no statistically significant difference between the performance of the AI-system and that of human experts with sensitivity, specificity, F1 and accuracy of 0.63, 0.78, 0.67 and 0.70 respectively.
Conclusion: This pilot study demonstrates the first multicenter application of an AI-based system in the prediction of submucosal invasion in endoscopic images of Barrett’s cancer. AI scored equal to international experts in the field, but more work is necessary to improve the system and apply it to video sequences and in a real-life setting. Nevertheless, the correct prediction of submucosal invasion in Barret´s cancer remains challenging for both experts and AI.
Limitations in computer-assisted diagnosis include lack of labeled data and inability to model the relation between what experts see and what computers learn. Even though artificial intelligence and machine learning have demonstrated remarkable performances in medical image computing, their accountability and transparency level must be improved to transfer this success into clinical practice. The reliability of machine learning decisions must be explained and interpreted, especially for supporting the medical diagnosis. While deep learning techniques are broad so that unseen information might help learn patterns of interest, human insights to describe objects of interest help in decision-making. This paper proposes a novel approach, DeepCraftFuse, to address the challenge of combining information provided by deep networks with visual-based features to significantly enhance the correct identification of cancerous tissues in patients affected with Barrett’s esophagus (BE). We demonstrate that DeepCraftFuse outperforms state-of-the-art techniques on private and public datasets, reaching results of around 95% when distinguishing patients affected by BE that is either positive or negative to esophageal cancer.
This work presents a systematic review concerning recent studies and technologies of machine learning for Barrett's esophagus (BE) diagnosis and treatment. The use of artificial intelligence is a brand new and promising way to evaluate such disease. We compile some works published at some well-established databases, such as Science Direct, IEEEXplore, PubMed, Plos One, Multidisciplinary Digital Publishing Institute (MDPI), Association for Computing Machinery (ACM), Springer, and Hindawi Publishing Corporation. Each selected work has been analyzed to present its objective, methodology, and results. The BE progression to dysplasia or adenocarcinoma shows a complex pattern to be detected during endoscopic surveillance. Therefore, it is valuable to assist its diagnosis and automatic identification using computer analysis. The evaluation of the BE dysplasia can be performed through manual or automated segmentation through machine learning techniques. Finally, in this survey, we reviewed recent studies focused on the automatic detection of the neoplastic region for classification purposes using machine learning methods.
The number of patients with Barret’s esophagus (BE) has increased in the last decades. Considering the dangerousness of the disease and its evolution to adenocarcinoma, an early diagnosis of BE may provide a high probability of cancer remission. However, limitations regarding traditional methods of detection and management of BE demand alternative solutions. As such, computer-aided tools have been recently used to assist in this problem, but the challenge still persists. To manage the problem, we introduce the infinity Restricted Boltzmann Machines (iRBMs) to the task of automatic identification of Barrett’s esophagus from endoscopic images of the lower esophagus. Moreover, since iRBM requires a proper selection of its meta-parameters, we also present a discriminative iRBM fine-tuning using six meta-heuristic optimization techniques. We showed that iRBMs are suitable for the context since it provides competitive results, as well as the meta-heuristic techniques showed to be appropriate for such task.
The development of adenocarcinoma in Barrett’s esophagus is difficult to detect by endoscopic surveillance of patients with signs of dysplasia. Computer assisted diagnosis of endoscopic images (CAD) could therefore be most helpful in the demarcation and classification of neoplastic lesions. In this study we tested the feasibility of a CAD method based on Speeded up Robust Feature Detection (SURF). A given database containing 100 images from 39 patients served as benchmark for feature based classification models. Half of the images had previously been diagnosed by five clinical experts as being ”cancerous”, the other half as ”non-cancerous”. Cancerous image regions had been visibly delineated (masked) by the clinicians. SURF features acquired from full images as well as from masked areas were utilized for the supervised training and testing of an SVM classifier. The predictive accuracy of the developed CAD system is illustrated by sensitivity and specificity values. The results based on full image matching where 0.78 (sensitivity) and 0.82 (specificity) were achieved, while the masked region approach generated results of 0.90 and 0.95, respectively.
Pixel-level classification is an essential part of computer vision. For learning from labeled data, many powerful deep learning models have been developed recently. In this work, we augment such supervised segmentation models by allowing them to learn from unlabeled data. Our semi-supervised approach, termed Error-Correcting Supervision, leverages a collaborative strategy. Apart from the supervised training on the labeled data, the segmentation network is judged by an additional network. The secondary correction network learns on the labeled data to optimally spot correct predictions, as well as to amend incorrect ones. As auxiliary regularization term, the corrector directly influences the supervised training of the segmentation network. On unlabeled data, the output of the correction network is essential to create a proxy for the unknown truth. The corrector’s output is combined with the segmentation network’s prediction to form the new target. We propose a loss function that incorporates both the pseudo-labels as well as the predictive certainty of the correction network. Our approach can easily be added to supervised segmentation models. We show consistent improvements over a supervised baseline on experiments on both the Pascal VOC 2012 and the Cityscapes datasets with varying amounts of labeled data.
In this work, we propose the use of single channel Color Co-occurrence Matrices for texture description of Barrett’sEsophagus (BE)and adenocarcinoma images. Further classification using supervised learning techniques, such as Optimum-Path Forest (OPF), Support Vector Machines with Radial Basisunction (SVM-RBF) and Bayesian classifier supports the contextof automatic BE and adenocarcinoma diagnosis. We validated three approaches of classification based on patches, patients and images in two datasets (MICCAI 2015 and Augsburg) using the color-and-texture descriptors and the machine learning techniques. Concerning MICCAI 2015 dataset, the best results were obtained using the blue channel for the descriptors and the supervised OPF for classification purposes in the patch-based approach, with sensitivity nearly to 73% for positive adenocarcinoma identification and specificity close to 77% for BE (non-cancerous) patch classification. Regarding the Augsburg dataset, the most accurate results were also obtained using both OPF classifier and blue channel descriptor for the feature extraction, with sensitivity close to 67% and specificity around to76%. Our work highlights new advances in the related research area and provides a promising technique that combines color and texture information, allied to three different approaches of dataset pre-processing aiming to configure robust scenarios for the classification step.
Computer-assisted analysis of endoscopic images can be helpful to the automatic diagnosis and classification of neoplastic lesions. Barrett's esophagus (BE) is a common type of reflux that is not straight forward to be detected by endoscopic surveillance, thus being way susceptible to erroneous diagnosis, which can cause cancer when not treated properly. In this work, we introduce the Optimum-Path Forest (OPF) classifier to the task of automatic identification of Barrett'sesophagus, with promising results and outperforming the well known Support Vector Machines (SVM) in the aforementioned context. We consider describing endoscopic images by means of feature extractors based on key point information, such as the Speeded up Robust Features (SURF) and Scale-Invariant Feature Transform (SIFT), for further designing a bag-of-visual-wordsthat is used to feed both OPF and SVM classifiers. The best results were obtained by means of the OPF classifier for both feature extractors, with values lying on 0.732 (SURF) - 0.735(SIFT) for sensitivity, 0.782 (SURF) - 0.806 (SIFT) for specificity, and 0.738 (SURF) - 0.732 (SIFT) for the accuracy.
Die Implantation eines Cochlea Implantates benötigt einen chirurgischen Zugang im Felsenbein und durch die Paukenhöhle des Patienten. Der Chirurg hat eine eingeschränkte Sicht im Operationsgebiet, die weiterhin viele Risikostrukturen enthält. Um eine Cochlea Implantation sicher und fehlerfrei durchzuführen, ist eine umfangreiche theoretische und praktische (teilweise berufsbegleitende) Fortbildung sowie langjährige Erfahrung notwendig. Unter Nutzung von realen klinischen CT/MRT Daten von Innen- und Mittelohr und der interaktiven Segmentierung der darin abgebildeten Strukturen (Nerven, Cochlea, Gehörknöchelchen,...) wird im HaptiVisT Projekt ein haptisch-visuelles Trainingssystem für die Implantation von Innen- und Mittelohr-Implantaten realisiert, das als sog. „Serious Game“ mit immersiver Didaktik gestaltet wird. Die Evaluierung des Demonstrators hinsichtlich Zweckmäßigkeit erfolgt prozessbegleitend und ergebnisorientiert, um mögliche technische oder didaktische Fehler vor Fertigstellung des Systems aufzudecken. Drei zeitlich versetzte Evaluationen fokussieren dabei chirurgisch-fachliche, didaktische sowie haptisch-ergonomische Akzeptanzkriterien.
Heavy smoke development represents an important challenge for operating physicians during laparoscopic procedures and can potentially affect the success of an intervention due to reduced visibility and orientation. Reliable and accurate recognition of smoke is therefore a prerequisite for the use of downstream systems such as automated smoke evacuation systems. Current approaches distinguish between non-smoked and smoked frames but often ignore the temporal context inherent in endoscopic video data. In this work, we therefore present a method that utilizes the pixel-wise displacement from randomly sampled images to the preceding frames determined using the optical flow algorithm by providing the transformed magnitude of the displacement as an additional input to the network. Further, we incorporate the temporal context at evaluation time by applying an exponential moving average on the estimated class probabilities of the model output to obtain more stable and robust results over time. We evaluate our method on two convolutional-based and one state-of-the-art transformer architecture and show improvements in the classification results over a baseline approach, regardless of the network used.
Training data for Neural Networks is often scarce in the medical domain, which often results in models that struggle to generalize and consequently showpoor performance on unseen datasets. Generally, adding augmentation methods to the training pipeline considerably enhances a model’s performance. Using the dataset of the Foot Ulcer Segmentation Challenge, we analyze two additional augmentation methods in the domain of chronic foot wounds - local warping of wound edges along with projection and blurring of shapes inside wounds. Our experiments show that improvements in the Dice similarity coefficient and Normalized Surface Distance metrics depend on a sensible selection of those augmentation methods.
Even though artificial intelligence and machine learning have demonstrated remarkable performances in medical image computing, their accountability and transparency level must be improved to transfer this success into clinical practice. The reliability of machine learning decisions must be explained and interpreted, especially for supporting the medical diagnosis.For this task, the deep learning techniques’ black-box nature must somehow be lightened up to clarify its promising results. Hence, we aim to investigate the impact of the ResNet-50 deep convolutional design for Barrett’s esophagus and adenocarcinoma classification. For such a task, and aiming at proposing a two-step learning technique, the output of each convolutional layer that composes the ResNet-50 architecture was trained and classified for further definition of layers that would provide more impact in the architecture. We showed that local information and high-dimensional features are essential to improve the classification for our task. Besides, we observed a significant improvement when the most discriminative layers expressed more impact in the training and classification of ResNet-50 for Barrett’s esophagus and adenocarcinoma classification, demonstrating that both human knowledge and computational processing may influence the correct learning of such a problem.