Zachow, Stefan
Refine
Year of publication
Document Type
- Article (94)
- In Proceedings (81)
- ZIB-Report (29)
- Book chapter (8)
- Poster (6)
- Master's Thesis (4)
- Other (4)
- Research data (4)
- Doctoral Thesis (2)
- In Collection (2)
Is part of the Bibliography
- no (235)
Keywords
- statistical shape and intensity models (5)
- Statistical shape analysis (3)
- 3d-reconstruction from 2d X-rays (2)
- 3d-reconstruction from 2d Xrays (2)
- Articulated Models, Statistical Shape And Intensity Models, 2D/3D Anatomy Reconstruction, Pelvic Parameters Measurement, Total Hip Arthroplasty (2)
- Classification (2)
- Kniegelenk (2)
- Lie groups (2)
- Manifold valued statistics (2)
- Medizinische Planung (2)
Institute
- Visual and Data-centric Computing (220)
- Visual Data Analysis (210)
- Therapy Planning (182)
- Numerical Mathematics (19)
- Computational Medicine (15)
- Visual Data Analysis in Science and Engineering (11)
- Geometric Data Analysis and Processing (6)
- Mathematics for Life and Materials Science (4)
- ZIB Allgemein (3)
- Bioinformatics in Medicine (1)
Motivation: The ever-rising volume of patients, high maintenance cost of operating rooms and time consuming analysis of surgical skills are fundamental problems that hamper the practical training of the next generation of surgeons. The hospitals prefer to keep the surgeons busy in real operations over training young surgeons for obvious economic reasons. One fundamental need in surgical training is the reduction of the time needed by the senior surgeon to review the endoscopic procedures performed by the young surgeon while minimizing the subjective bias in evaluation. The unprecedented performance of deep learning ushers the new age of data-driven automatic analysis of surgical skills.
Method: Deep learning is capable of efficiently analyzing thousands of hours of laparoscopic video footage to provide an objective assessment of surgical skills. However, the traditional end-to-end setting of deep learning (video in, skill assessment out) is not explainable. Our strategy is to utilize the surgical process modeling framework to divide the surgical process into understandable components. This provides the opportunity to employ deep learning for superior yet automatic detection and evaluation of several aspects of laparoscopic cholecystectomy such as surgical tool and phase detection.
We employ ZIBNet for the detection of surgical tool presence. ZIBNet employs pre-processing based on tool usage imbalance, a transfer learned 50-layer residual network (ResNet-50) and temporal smoothing. To encode the temporal evolution of tool usage (over the entire video sequence) that relates to the surgical phases, Long Short Term Memory (LSTM) units are employed with long-term dependency.
Dataset: We used CHOLEC 80 dataset that consists of 80 videos of laparoscopic cholecystectomy performed by 13 surgeons, divided equally for training and testing. In these videos, up to three different tools (among 7 types of tools) can be present in a frame.
Results: The mean average precision of the detection of all tools is 93.5 ranging between 86.8 and 99.3, a significant improvement (p <0.01) over the previous state-of-the-art. We observed that less frequent tools like Scissors, Irrigator, Specimen Bag etc. are more related to phase transitions. The overall precision (recall) of the detection of all surgical phases is 79.6 (81.3).
Conclusion: While this is not the end goal for surgical skill analysis, the development of such a technological platform is essential toward a data-driven objective understanding of surgical skills. In future, we plan to investigate surgeon-in-the-loop analysis and feedback for surgical skill analysis.
Automatic recognition of surgical phases is an important component for developing an intra-operative context-aware system. Prior work in this area focuses on recognizing short-term tool usage patterns within surgical phases. However, the difference between intra- and inter-phase tool usage patterns has not been investigated for automatic phase recognition. We developed a Recurrent Neural Network (RNN), in particular a state-preserving Long Short Term Memory (LSTM) architecture to utilize the long-term evolution of tool usage within complete surgical procedures. For fully automatic tool presence detection from surgical video frames, a Convolutional Neural Network (CNN) based architecture namely ZIBNet is employed. Our proposed approach outperformed EndoNet by 8.1% on overall precision for phase detection tasks and 12.5% on meanAP for tool recognition tasks.
Surgical tool segmentation in endoscopic videos is an important component of computer assisted interventions systems. Recent success of image-based solutions using fully-supervised deep learning approaches can be attributed to the collection of big labeled datasets. However, the annotation of a big dataset of real videos can be prohibitively expensive and time consuming. Computer simulations could alleviate the manual labeling problem, however, models trained on simulated data do not generalize to real data. This work proposes a consistency-based framework for joint learning of simulated and real (unlabeled) endoscopic data to bridge this performance generalization issue. Empirical results on two data sets (15 videos of the Cholec80 and EndoVis'15 dataset) highlight the effectiveness of the proposed Endo-Sim2Real method for instrument segmentation. We compare the segmentation of the proposed approach with state-of-the-art solutions and show that our method improves segmentation both in terms of quality and quantity.
This study’s objective was the generation of a standardized geometry of the healthy nasal cavity.
An average geometry of the healthy nasal cavity was generated using a statistical shape model based on 25 symptom-free subjects. Airflow within the average geometry and these geometries was calculated using fluid simulations. Integral measures of the nasal resistance, wall shear stresses (WSS) and velocities were calculated as well as cross-sectional areas (CSA). Furthermore, individual WSS and static pressure distributions were mapped onto the average geometry.
The average geometry featured an overall more regular shape that resulted in less resistance, reduced wall shear stresses and velocities compared to the median of the 25 geometries. Spatial distributions of WSS and pressure of average geometry agreed well compared to the average distributions of all individual geometries. The minimal CSA of the average geometry was larger than the median of all individual geometries (83.4 vs. 74.7 mm²).
The airflow observed within the average geometry of the healthy nasal cavity did not equal the average airflow of the individual geometries. While differences observed for integral measures were notable, the calculated values for the average geometry lay within the distributions of the individual parameters. Spatially resolved parameters differed less prominently.
We describe a novel nonlinear statistical shape model basedon differential coordinates viewed as elements of GL+(3). We adopt an as-invariant-as possible framework comprising a bi-invariant Lie group mean and a tangent principal component analysis based on a unique GL+(3)-left-invariant, O(3)-right-invariant metric. Contrary to earlier work that equips the coordinates with a specifically constructed group structure, our method employs the inherent geometric structure of the group-valued data and therefore features an improved statistical power in identifying shape differences. We demonstrate this in experiments on two anatomical datasets including comparison to the standard Euclidean as well as recent state-of-the-art nonlinear approaches to statistical shape modeling.
We present a novel approach for nonlinear statistical shape modeling that is invariant under Euclidean motion and thus alignment-free. By analyzing metric distortion and curvature of shapes as elements of Lie groups in a consistent Riemannian setting, we construct a framework that reliably handles large deformations. Due to the explicit character of Lie group operations, our non-Euclidean method is very efficient allowing for fast and numerically robust processing. This facilitates Riemannian analysis of large shape populations accessible through longitudinal and multi-site imaging studies providing increased statistical power. We evaluate the performance of our model w.r.t. shape-based classification of pathological malformations of the human knee and show that it outperforms the standard Euclidean as well as a recent nonlinear approach especially in presence of sparse training data. To provide insight into the model's ability of capturing natural biological shape variability, we carry out an analysis of specificity and generalization ability.
We describe a novel nonlinear statistical shape model basedon differential coordinates viewed as elements of GL+(3). We adopt an as-invariant-as possible framework comprising a bi-invariant Lie group mean and a tangent principal component analysis based on a unique GL+(3)-left-invariant, O(3)-right-invariant metric. Contrary to earlier work that equips the coordinates with a specifically constructed group structure, our method employs the inherent geometric structure of the group-valued data and therefore features an improved statistical power in identifying shape differences. We demonstrate this in experiments on two anatomical datasets including comparison to the standard Euclidean as well as recent state-of-the-art nonlinear approaches to statistical shape modeling.
Quantification of magnetic resonance (MR)-based relaxation parameters of tendons and ligaments is challenging due to their very short transverse relaxation times, requiring application of ultra-short echo-time (UTE) imaging sequences. We quantify both T1 and T2⁎ in the quadriceps and patellar tendons of healthy volunteers at a field strength of 3 T and visualize the results based on 3D segmentation by using bivariate histogram analysis. We applied a 3D ultra-short echo-time imaging sequence with either variable repetition times (VTR) or variable flip angles (VFA) for T1 quantification in combination with multi-echo acquisition for extracting T2⁎. The values of both relaxation parameters were subsequently binned for bivariate histogram analysis and corresponding cluster identification, which were subsequently visualized. Based on manually-drawn regions of interest in the tendons on the relaxation parameter maps, T1 and T2⁎ boundaries were selected in the bivariate histogram to segment the quadriceps and patellar tendons and visualize the relaxation times by 3D volumetric rendering. Segmentation of bone marrow, fat, muscle and tendons was successfully performed based on the bivariate histogram analysis. Based on the segmentation results mean T2⁎ relaxation times, over the entire tendon volumes averaged over all subjects, were 1.8 ms ± 0.1 ms and 1.4 ms ± 0.2 ms for the patellar and quadriceps tendons, respectively. The mean T1 value of the patellar tendon, averaged over all subjects, was 527 ms ± 42 ms and 476 ms ± 40 ms for the VFA and VTR acquisitions, respectively. The quadriceps tendon had higher mean T1 values of 662 ms ± 97 ms (VFA method) and 637 ms ± 40 ms (VTR method) compared to the patellar tendon. 3D volumetric visualization of the relaxation times revealed that T1 values are not constant over the volume of both tendons, but vary locally. This work provided additional data to build upon the scarce literature available on relaxation times in the quadriceps and patellar tendons. We were able to segment both tendons and to visualize the relaxation parameter distributions over the entire tendon volumes.
In dentistry, software-based medical image analysis and visualization provide efficient and accurate diagnostic and therapy planning capabilities. We present an approach for the automatic recognition of tooth types and positions in digital volume tomography (DVT). By using deep learning techniques in combination with dimensionality reduction through non-planar reformatting of the jaw anatomy, DVT data can be efficiently processed and
teeth reliably recognized and classified, even in the presence of imaging artefacts, missing or dislocated teeth.
We evaluated our approach, which is based on 2D Convolutional Neural Networks (CNNs), on 118 manually
annotated cases of clinical DVT datasets. Our proposed method correctly classifies teeth with an accuracy of
94% within a limit of 2mm distance to ground truth labels.
Statistical Shape Models (SSMs) are a proven means for model-based 3D anatomy reconstruction from medical
image data. In orthopaedics and biomechanics, SSMs are increasingly employed to individualize measurement
data or to create individualized anatomical models to which implants can be adapted to or functional tests can be performed on. For modeling and analysis of articulated structures, so called articulated SSMs (aSSMs) have been developed. However, a missing feature of aSSMs is the consideration of collisions in the course of individual fitting and articulation. The aim of our work was to develop aSSMs that handle collisions between components correctly. That way it becomes possible to adjust shape and articulation in view of a physically and geometrically plausible individualization. To be able to apply collision-aware aSSMs in simulation and optimisation, our approach is based on an ecient collision detection method employing Graphics Processing Units (GPUs).