Effective Disjoint Representational Learning for Anatomical Segmentation

Abstract

In the wake of the limited availability of pertinent datasets, the application of computer vision methods for semantic segmentation of abdominal structures is mainly constrained to surgical instruments or organ-specific segmentations. Multi-organ segmentation has the potential to furnish supplementary assistance in multifarious domains in healthcare, for instance, robot-assisted laparoscopic surgery. However, in addition to the complexity involved in discriminating anatomical structures due to their visual attributes and operative conditions, the representation bias pertaining to organ size results in poor segmentation performance on organs with smaller pixel proportions. In this work, we focus on alleviating the influence of representation bias by involving different encoder-decoder frameworks for learning organ-specific features. In particular, we investigate the effect of organ-specific decoders on binary segmentation of anatomical structures in abdominal surgery. Additionally, we analyze the effect of organ-specific pretraining on the multi-label segmentation in two model training settings including knowledge sharing and disjoint learning, in relation to the contextual feature sharing between organ-specific decoders. Our results illustrate the significant gain in segmentation performance by incorporating organ-specific decoders, especially for less represented organs.