Clinical priors-inspired privileged knowledge distillation for reliable pancreatic lesion classification.
Authors
Affiliations (7)
Affiliations (7)
- Department of Gastroenterology, Center for Leading Medicine, Division of Life Sciences and Medicine, Advanced Technologies of IHM, the First Affiliated Hospital of USTC, University of Science and Technology of China, Hefei, 230001, China.
- School of Information Science and Technology, University of Science and Technology of China, Hefei, 230027, China.
- Department of Gastroenterology, the First Affiliated Hospital of Anhui Medical University, Hefei, 230022, China.
- Department of Gastroenterology, the Second Affiliated Hospital of Anhui Medical University, Hefei, 230601, China.
- Department of Gastroenterology, the First People's Hospital of Hefei, Hefei, 230061, China.
- School of Information Science and Technology, University of Science and Technology of China, Hefei, 230027, China. Electronic address: [email protected].
- Department of Gastroenterology, Center for Leading Medicine, Division of Life Sciences and Medicine, Advanced Technologies of IHM, the First Affiliated Hospital of USTC, University of Science and Technology of China, Hefei, 230001, China. Electronic address: [email protected].
Abstract
Accurate classification of pancreatic lesions is critical for guiding treatment decisions, with computed tomography (CT) being the primary modality owing to its high spatial resolution. Although deep learning has advanced CT-based pancreatic lesion classification, its inherently low soft-tissue contrast complicates the diagnosis of challenging samples, such as small or isoattenuating lesions. Multi-modal learning is a promising alternative but is often clinically impractical, as it requires all modalities during inference and fails to fully leverage the clinically relevant complementary information. To address these limitations, we propose a customized Clinical priors-inspired Privileged Knowledge Distillation (CPKD) framework that employs magnetic resonance imaging (MRI) as a privileged modality for reliable pancreatic lesion classification. It introduces two novel components guided by clinical priors: (1) a Diagnosis-related Privileged Knowledge (DPK) strategy, which captures radiological features based on auxiliary tasks, distilling them into the primary CT branch via dual-level distillation. (2) a Prior-guided Dual Calibration (PDC) strategy, which rectifies biologically implausible predictions and adaptively strengthens knowledge transfer for the most challenging samples by integrating both biological prediction calibration and personalized distillation calibration. Extensive experiments on a multi-center dataset demonstrate that CPKD consistently outperforms other competing methods. Most notably, it achieves an accuracy gain of 8.2% over the second-best method in a challenging subgroup of small or isoattenuating lesions. These findings highlight that by integrating clinical priors into the knowledge distillation process, CPKD provides a more reliable and clinically plausible solution for pancreatic lesion classification. Code is available at https://github.com/Hanqiaoyu/CPKD.