Regensburg Medical Image Computing (ReMIC)
Refine
Document Type
Is part of the Bibliography
- no (37)
Keywords
- Diagnose (12)
- Künstliche Intelligenz (11)
- Maschinelles Lernen (10)
- Artificial Intelligence (9)
- Speiseröhrenkrebs (7)
- Machine learning (6)
- Speiseröhrenkrankheit (6)
- Adenocarcinoma (5)
- Barrett's esophagus (5)
- Deep Learning (5)
Institute
Begutachtungsstatus
- peer-reviewed (32)
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.
Background
This study evaluated the effect of an artificial intelligence (AI)-based clinical decision support system on the performance and diagnostic confidence of endoscopists in their assessment of Barrett’s esophagus (BE).
Methods
96 standardized endoscopy videos were assessed by 22 endoscopists with varying degrees of BE experience from 12 centers. Assessment was randomized into two video sets: group A (review first without AI and second with AI) and group B (review first with AI and second without AI). Endoscopists were required to evaluate each video for the presence of Barrett’s esophagus-related neoplasia (BERN) and then decide on a spot for a targeted biopsy. After the second assessment, they were allowed to change their clinical decision and confidence level.
Results
AI had a stand-alone sensitivity, specificity, and accuracy of 92.2%, 68.9%, and 81.3%, respectively. Without AI, BE experts had an overall sensitivity, specificity, and accuracy of 83.3%, 58.1%, and 71.5%, respectively. With AI, BE nonexperts showed a significant improvement in sensitivity and specificity when videos were assessed a second time with AI (sensitivity 69.8% [95%CI 65.2%–74.2%] to 78.0% [95%CI 74.0%–82.0%]; specificity 67.3% [95%CI 62.5%–72.2%] to 72.7% [95%CI 68.2%–77.3%]). In addition, the diagnostic confidence of BE nonexperts improved significantly with AI.
Conclusion
BE nonexperts benefitted significantly from additional AI. BE experts and nonexperts remained significantly below the stand-alone performance of AI, suggesting that there may be other factors influencing endoscopists’ decisions to follow or discard AI advice.
Aims
Recent evidence suggests the possibility of intraprocedural phase recognition in surgical operations as well as endoscopic interventions such as peroral endoscopic myotomy and endoscopic submucosal dissection (ESD) by AI-algorithms. The intricate measurement of intraprocedural phase distribution may deepen the understanding of the procedure. Furthermore, real-time quality assessment as well as automation of reporting may become possible. Therefore, we aimed to develop an AI-algorithm for intraprocedural phase recognition during ESD.
Methods
A training dataset of 364385 single images from 9 full-length ESD videos was compiled. Each frame was classified into one procedural phase. Phases included scope manipulation, marking, injection, application of electrical current and bleeding. Allocation of each frame was only possible to one category. This training dataset was used to train a Video Swin transformer to recognize the phases. Temporal information was included via logarithmic frame sampling. Validation was performed using two separate ESD videos with 29801 single frames.
Results
The validation yielded sensitivities of 97.81%, 97.83%, 95.53%, 85.01% and 87.55% for scope manipulation, marking, injection, electric application and bleeding, respectively. Specificities of 77.78%, 90.91%, 95.91%, 93.65% and 84.76% were measured for the same parameters.
Conclusions
The developed algorithm was able to classify full-length ESD videos on a frame-by-frame basis into the predefined classes with high sensitivities and specificities. Future research will aim at the development of quality metrics based on single-operator phase distribution.
Aims
While AI has been successfully implemented in detecting and characterizing colonic polyps, its role in therapeutic endoscopy remains to be elucidated. Especially third space endoscopy procedures like ESD and peroral endoscopic myotomy (POEM) pose a technical challenge and the risk of operator-dependent complications like intraprocedural bleeding and perforation. Therefore, we aimed at developing an AI-algorithm for intraprocedural real time vessel detection during ESD and POEM.
Methods
A training dataset consisting of 5470 annotated still images from 59 full-length videos (47 ESD, 12 POEM) and 179681 unlabeled images was used to train a DeepLabV3+neural network with the ECMT semi-supervised learning method. Evaluation for vessel detection rate (VDR) and time (VDT) of 19 endoscopists with and without AI-support was performed using a testing dataset of 101 standardized video clips with 200 predefined blood vessels. Endoscopists were stratified into trainees and experts in third space endoscopy.
Results
The AI algorithm had a mean VDR of 93.5% and a median VDT of 0.32 seconds. AI support was associated with a statistically significant increase in VDR from 54.9% to 73.0% and from 59.0% to 74.1% for trainees and experts, respectively. VDT significantly decreased from 7.21 sec to 5.09 sec for trainees and from 6.10 sec to 5.38 sec for experts in the AI-support group. False positive (FP) readings occurred in 4.5% of frames. FP structures were detected significantly shorter than true positives (0.71 sec vs. 5.99 sec).
Conclusions
AI improved VDR and VDT of trainees and experts in third space endoscopy and may reduce performance variability during training. Further research is needed to evaluate the clinical impact of this new technology.
Aims
Artificial Intelligence (AI) systems in gastrointestinal endoscopy are narrow because they are trained to solve only one specific task. Unlike Narrow-AI, general AI systems may be able to solve multiple and unrelated tasks. We aimed to understand whether an AI system trained to detect, characterize, and segment early Barrett’s neoplasia (Barrett’s AI) is only capable of detecting this pathology or can also detect and segment other diseases like early squamous cell cancer (SCC).
Methods
120 white light (WL) and narrow-band endoscopic images (NBI) from 60 patients (1 WL and 1 NBI image per patient) were extracted from the endoscopic database of the University Hospital Augsburg. Images were annotated by three expert endoscopists with extensive experience in the diagnosis and endoscopic resection of early esophageal neoplasias. An AI system based on DeepLabV3+architecture dedicated to early Barrett’s neoplasia was tested on these images. The AI system was neither trained with SCC images nor had it seen the test images prior to evaluation. The overlap between the three expert annotations („expert-agreement“) was the ground truth for evaluating AI performance.
Results
Barrett’s AI detected early SCC with a mean intersection over reference (IoR) of 92% when at least 1 pixel of the AI prediction overlapped with the expert-agreement. When the threshold was increased to 5%, 10%, and 20% overlap with the expert-agreement, the IoR was 88%, 85% and 82%, respectively. The mean Intersection Over Union (IoU) – a metric according to segmentation quality between the AI prediction and the expert-agreement – was 0.45. The mean expert IoU as a measure of agreement between the three experts was 0.60.
Conclusions
In the context of this pilot study, the predictions of SCC by a Barrett’s dedicated AI showed some overlap to the expert-agreement. Therefore, features learned from Barrett’s cancer-related training might be helpful also for SCC prediction. Our results allow different possible explanations. On the one hand, some Barrett’s cancer features generalize toward the related task of assessing early SCC. On the other hand, the Barrett’s AI is less specific to Barrett’s cancer than a general predictor of pathological tissue. However, we expect to enhance the detection quality significantly by extending the training to SCC-specific data. The insight of this study opens the way towards a transfer learning approach for more efficient training of AI to solve tasks in other domains.
Aims
Endoscopic retrograde cholangiopancreaticography (ERCP) is the gold standard in the diagnosis as well as treatment of diseases of the pancreatobiliary tract. However, it is technically complex and has a relatively high complication rate. In particular, cannulation of the papillary ostium remains challenging. The aim of this study is to examine whether a deep-learning algorithm can be used to detect the major duodenal papilla and in particular the papillary ostium reliably and could therefore be a valuable tool for inexperienced endoscopists, particularly in training situation.
Methods
We analyzed a total of 654 retrospectively collected images of 85 patients. Both the major duodenal papilla and the ostium were then segmented. Afterwards, a neural network was trained using a deep-learning algorithm. A 5-fold cross-validation was performed. Subsequently, we ran the algorithm on 5 prospectively collected videos of ERCPs.
Results
5-fold cross-validation on the 654 labeled data resulted in an F1 value of 0.8007, a sensitivity of 0.8409 and a specificity of 0.9757 for the class papilla, and an F1 value of 0.5724, a sensitivity of 0.5456 and a specificity of 0.9966 for the class ostium. Regardless of the class, the average F1 value (class papilla and class ostium) was 0.6866, the sensitivity 0.6933 and the specificity 0.9861. In 100% of cases the AI-detected localization of the papillary ostium in the prospectively collected videos corresponded to the localization of the cannulation performed by the endoscopist.
Conclusions
In the present study, the neural network was able to identify the major duodenal papilla with a high sensitivity and high specificity. In detecting the papillary ostium, the sensitivity was notably lower. However, when used on videos, the AI was able to identify the location of the subsequent cannulation with 100% accuracy. In the future, the neural network will be trained with more data. Thus, a suitable tool for ERCP could be established, especially in the training situation.
Effect of AI on performance of endoscopists to detect Barrett neoplasia: A Randomized Tandem Trial
()
Background and study aims
To evaluate the effect of an AI-based clinical decision support system (AI) on the performance and diagnostic confidence of endoscopists during the assessment of Barrett's esophagus (BE).
Patients and Methods
Ninety-six standardized endoscopy videos were assessed by 22 endoscopists from 12 different centers with varying degrees of BE experience.
The assessment was randomized into two video sets: Group A (review first without AI and second with AI) and group B (review first with AI and second without AI). Endoscopists were required to evaluate each video for the presence of Barrett's esophagus-related neoplasia (BERN) and then decide on a spot for a targeted biopsy. After the second assessment, they were allowed to change their clinical decision and confidence level.
Results
AI had a standalone sensitivity, specificity, and accuracy of 92.2%, 68.9%, and 81.6%, respectively. Without AI, BE experts had an overall sensitivity, specificity, and accuracy of 83.3%, 58.1 and 71.5%, respectively. With AI, BE nonexperts showed a significant improvement in sensitivity and specificity when videos were assessed a second time with AI (sensitivity 69.7% (95% CI, 65.2% - 74.2%) to 78.0% (95% CI, 74.0% - 82.0%); specificity 67.3% (95% CI, 62.5% - 72.2%) to 72.7% (95 CI, 68.2% - 77.3%). In addition, the diagnostic confidence of BE nonexperts improved significantly with AI.
Conclusion
BE nonexperts benefitted significantly from the additional AI. BE experts and nonexperts remained below the standalone performance of AI, suggesting that there may be other factors influencing endoscopists to follow or discard AI advice.
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.
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.
Einleitung Die Endoskopische Retrograde Cholangiopankreatikographie (ERCP) ist der Goldstandard in der Diagnostik und Therapie von Erkrankungen des pankreatobiliären Trakts. Jedoch ist sie technisch sehr anspruchsvoll und weist eine vergleichsweise hohe Komplikationsrate auf.
Ziele
In der vorliegenden Machbarkeitsstudie soll geprüft werden, ob mithilfe eines Deep-learning-Algorithmus die Papille und das Ostium zuverlässig detektiert werden können und somit für Endoskopiker mit geringer Erfahrung ein geeignetes Hilfsmittel, insbesondere für die Ausbildungssituation, darstellen könnten.
Methodik
Wir betrachteten insgesamt 606 Bilddatensätze von 65 Patienten. In diesen wurde sowohl die Papilla duodeni major als auch das Ostium segmentiert. Anschließend wurde eine neuronales Netz mittels eines Deep-learning-Algorithmus trainiert. Außerdem erfolgte eine 5-fache Kreuzvaldierung.
Ergebnisse
Bei einer 5-fachen Kreuzvaldierung auf den 606 gelabelten Daten konnte für die Klasse Papille eine F1-Wert von 0,7908, eine Sensitivität von 0,7943 und eine Spezifität von 0,9785 erreicht werden, für die Klasse Ostium eine F1-Wert von 0,5538, eine Sensitivität von 0,5094 und eine Spezifität von 0,9970 (vgl. [Tab. 1]). Unabhängig von der Klasse zeigte sich gemittelt (Klasse Papille und Klasse Ostium) ein F1-Wert von 0,6673, eine Sensitivität von 0,6519 und eine Spezifität von 0,9877 (vgl. [Tab. 2]).
Schlussfolgerung
In vorliegende Machbarkeitsstudie konnte das neuronale Netz die Papilla duodeni major mit einer hohen Sensitivität und sehr hohen Spezifität identifizieren. Bei der Detektion des Ostiums war die Sensitivität deutlich geringer. Zukünftig soll das das neuronale Netz mit mehr Daten trainiert werden. Außerdem ist geplant, den Algorithmus auch auf Videos anzuwenden. Somit könnte langfristig ein geeignetes Hilfsmittel für die ERCP etabliert werden.