Cross-modal contrastive knowledge distillation with hierarchical decomposition for kidney tumor segmentation in non-contrast CT

Kidney cancer is one of the most prevalent malignancies worldwide, often remaining asymptomatic in its early stages, leading to frequent incidental detection during routine CT screenings. While contrast-enhanced CT (CECT) provides rich contrast information that facilitates tumor delineation, non-contrast CT (NCCT) is more commonly used in routine screenings and for patients who cannot receive contrast agents. However, NCCT suffers from low contrast differences between tissues, making kidney tumor segmentation extremely challenging for existing deep learning models. To address this issue, we propose a novel Cross-Modal Contrastive Knowledge Distillation with Hierarchical Decomposition framework for NCCT-based kidney tumor segmentation. Our framework utilizes paired CECT and NCCT data during training to transfer rich contrast information from CECT to NCCT, while requiring only NCCT images during inference. To achieve more effective contrast feature transfer, we hierarchically decompose the segmentation task by separating kidney structure knowledge and contrast knowledge for kidney-tumor differentiation. This allows us to explicitly distill contrast-related knowledge from CECT into NCCT, focusing on the features required for accurate tumor segmentation. Furthermore, we introduce a difficulty-aware cross-modal contrastive learning strategy that aligns contrast-deficient NCCT tumor features with contrast-rich tumor representations from CECT while prioritizing challenging NCCT tumor regions that are more likely to be misclassified. Extensive experiments on kidney tumor datasets demonstrate that our framework significantly outperforms state-of-the-art segmentation methods, achieving superior accuracy on NCCT by leveraging contrast knowledge from CECT during training. These results highlight the clinical applicability of our approach in routine CT screening settings, providing a robust non-contrast CT-based tumor segmentation model that benefits from contrast-rich knowledge transfer.