Exploring Explainable Optimization in Medical Segmentation Network for Multi-Scale Generalization with Anatomical Atlas

Abstract

Significant efforts have been directed toward integrating machine learning into visual recognition tasks in medical domain, especially for medical image segmentation. Image segmentation provides pixel/voxel-wise localization of each organ/tissue target and allows researchers and clinicians to perform quantitative measures for investigating biomarkers. However, annotating volumetric labels is time-consuming. There is limited interpretability of features across learned current models. Furthermore, the volumetric morphology of organ of interests significantly varies with the substantial variability of demographics across population. Such variability limits the feasibility of generalizing population-wise features on specific organs to further investigate the corresponding biomarkers in conditions. In this dissertation, we first investigate training strategies to enhance the robustness in deep learning models across multi-modality imaging. We were inspired by the current progress of vision transformers and revisited the capabilities of large kernel convolution. We model the spatial frequency of the human visual behavior and derive a theoretical re-parameterization strategy to enhance both the learning capability and explainability of large kernel convolution for volumetric segmentation. With the basis of current network designs, we investigate self-supervised learning strategy to dynamically adapt multi-contrast imaging and integrate multiple semantic meanings to enhance the feature explainabilty, thus enhancing model generalizability with interpretable features. With the generated segmentation across large population cohort, we create multiple atlas templates to generalize the population characteristics across organs, and thus to facilitate the progress of revealing distinctive organ-specific biomarkers and align multi-scale findings from cellular level to organ level.