When evidence modeling meets knowledge distillation: Towards reliable contrast-enhanced knowledge distillation for...

Contrast-enhanced knowledge distillation promises to transform medical diagnostics and reveal promising approaches for tumor segmentation on non-contrast medical images. However, existing methods related to contrast-enhanced knowledge distillation still make it hard to distill reliable contrast-enhanced knowledge for tumor segmentation due to the limitations of (1) unable to quantify uncertainty information for reliable contrast-enhanced and non-contrast knowledge modeling, which leads to an over-confidence cross-domain adaptation for transferring contrast-enhanced knowledge; (2) using vision information only ignores rich semantic features in medical language, which make it hard to model complex tumor enhancement feature. In this study, we propose an evidence-guided and tumor-aware knowledge distillation (EGTA-KD) for transferring contrast-enhanced domain knowledge to non-contrast domain knowledge. Specifically, to achieve tumor-awareness by embedding semantic features from text, the tumor-aware cross-modal synchronizer (TACMS) is proposed to calculate tumor score maps for matching pixel wise image and text features. To achieve reliable cross-domain modeling for transferring contrast-enhanced knowledge, the innovative uncertainty-quantified evidence unit (UQEU) parameterizes the probability distribution within subjective logic to gather reliable evidence of contrast-enhanced knowledge while quantifying the uncertainty of prediction. Lastly, newly designed dual-level knowledge distillation (DLKD) minimizes tumor score map errors and matches evidence distribution for uncertainty-aware contrast-enhanced knowledge distillation. Extensive experiments of tumor segmentation on non-contrast medical images are performed using multi-modality medical image datasets (i.e., Brain MRI dataset, Liver MRI dataset, and Kidney CT dataset). Experimental results demonstrate the proposed EGTA-KD outperforms the other compared state-of-the-art methods, revealing its superiority of tumor segmentation on non-contrast medical images via uncertainty-aware contrast-enhanced knowledge distillation.