Multi-organ Abdominal Segmentation with Mesh-Based Bayesian Neural Networks

• Applicant: Professor Dr. Christian Wachinger
• Subject Area: Medical Informatics and Medical Bioinformatics ; Epidemiology and Medical Biometry/Statistics ; Radiology
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 460880779

Ongoing population studies will provide an unprecedented resource of magnetic resonance imaging (MRI) of the abdomen, which may yield novel scientific discoveries and the identification of imaging biomarkers. However, the data analysis on such a large scale will require novel, fully-automated approaches. A crucial first step for many analysis tasks is the segmentation of organs, which is very time-consuming when performed manually. Currently, software tools do not exist for the fully automated and robust segmentation of abdominal organs in MRI. The automated segmentation is a challenging task because of imaging artifacts, limited resolution, and anatomical variability. To address these issues, we will develop a novel approach based on recent advances in deep learning. The objective of this application is to develop a mesh-based deep neural network for the segmentation of abdominal organs in MRI population studies. The seven organs we will segment are: liver, pancreas, spleen, left/right kidney, and left/right adrenal gland. Our method will directly estimate the organ mesh without the need of an independent mesh creation step, which will enhance follow-up analyses. Furthermore, having an explicit contour will enable us to enforce additional geometrical constraints during training. Of particular interest will be the design of a highly robust network, which we aim to achieve by enforcing segmentations to be concordant with prior knowledge about the abdominal anatomy. Moreover, we want to achieve consistent segmentations across datasets by using adversarial training. We will further design a new approach for automated quality control of segmentations by estimating the uncertainty of the segmentation with a Bayesian approach. In contrast to prior work, this we will not estimate uncertainty per voxel but per vertex of the mesh. Finally, we will include additional technical innovations to fit the network in the GPU memory and to train the network with a limited number of annotated images. The developed method is intended for the large-scale analysis of population data and will enable researchers to study properties of abdominal organs with the wealth of other data available in these studies. The key benefits of our approach are: (i) robustness, which will provide reliable segmentations without manual interactions, (ii) consistency, which will enable the joint analysis across population studies, (iii) automated quality control, which will facilitate the follow-up analysis without the need for manually reviewing the segmentation masks. The project will advance the state-of-the-art in medical image segmentation and allow for novel avenues in analyzing abdominal organs in population studies.

• DFG Programme: Research Grants