Adaptive Knowledge Distillation for Anatomical Segmentation in Pelvic CT Imaging of Prostate Cancer

Accurate delineation of the prostate and surrounding pelvic structures is critical to successful treatment planning, accurate identification, and staging of prostate cancer. Segmentation of anatomically complex regions surrounding the prostate in CT imaging can be challenging due to low soft-tissue contrast and complex boundary delineations. In this work, we investigate three complementary paradigms of knowledge distillation-voxel-level, region-level, and a dynamically weighted combination of the two-to improve segmentation performance for the prostate and parailiac regions.

The region-level approach imposes the semantic coherence of network predictions via the region-wise contrastive form of supervision, whereas the voxel-level distillation provides fine-tuned supervision in terms of Kullback-Leibler divergence on soft probabilistic outputs. We introduce a novel fusion approach that adds uncertainty-aware dynamic weighting, thus allowing the model to adjust the contribution of every distillation loss in an adaptive manner during training, taking advantage of the strengths of both approaches. The distillation methods are implemented within the dual-network architecture in terms of VNet (Milletari et al. in Proc. Int. Conf. 3D Vis. (3DV):565-571, 2016) and 3D-ResVNet (Wang et al. in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR)), thus allowing synergistic learning of different architectural biases.

Experimental results on both an in-house collected and annotated CT dataset of prostate cancer patients-where parailiac regions and the prostate gland (including seminal vesicles) are manually segmented-and three public benchmark datasets demonstrate that each individual distillation method consistently improves segmentation accuracy over baseline models. These results indicate that the effectiveness of the proposed distillation strategies generalizes across different datasets and anatomical structures, highlighting their robustness and practical applicability.

Complementary supervision at voxel and region levels can improve the delineation of complex pelvic structures in CT imaging of prostate cancer.