Scholarlink

<p>Infectious pancreatic necrosis (IPN) is a severe complication of acute necrotizing pancreatitis with high morbidity and mortality. Early identification from computed tomography (CT) is clinically important but remains challenging due to subtle imaging manifestations, the lack of lesion-level annotations, and heterogeneous disease severity. Although deep learning methods have shown promise for CT-based IPN prediction, most existing approaches underutilize severity-related information and provide limited spatial interpretability, restricting their robustness and clinical applicability. In this work, we propose a severity-consistent weakly supervised framework for CT-based IPN prediction with class activation mapping (CAM)-guided lesion localization. The framework leverages routinely available heterogeneous clinical annotations, including scan-level IPN labels and the computed tomography severity index (CTSI), without requiring pixel-level infection annotations. A segmentation-guided multi-task learning strategy is employed to jointly model IPN prediction and CTSI estimation using pancreas-centered representations, thereby stabilizing feature learning under weak supervision. To incorporate clinical severity knowledge, a ranking-based severity consistency regularization is introduced to encourage a monotonic relationship between disease severity and predicted IPN risk. In addition, a classification-to-localization weak supervision mechanism is developed, in which anatomically constrained CAMs are distilled into lesion representations and reintegrated into IPN prediction through lesion-aware pooling. Experimental results demonstrate that the proposed framework substantially improves IPN prediction performance, achieving an area under the receiver operating characteristic curve of 0.842 with a sensitivity of 0.796, while simultaneously providing visually plausible lesion localization. These findings suggest that the proposed approach offers a practical and interpretable solution for CT-based IPN risk assessment under limited annotation settings.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A weakly supervised framework for CT-based infectious pancreatic necrosis prediction with CAM-guided lesion localization

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Ziyao Meng,
  • Jiajia Li,
  • Yuechen Liu,
  • Hao Shen,
  • Ziwei Zhao,
  • Yunfan Liu,
  • Jie Dong,
  • Haitao Song,
  • Nan Jiang

摘要

Infectious pancreatic necrosis (IPN) is a severe complication of acute necrotizing pancreatitis with high morbidity and mortality. Early identification from computed tomography (CT) is clinically important but remains challenging due to subtle imaging manifestations, the lack of lesion-level annotations, and heterogeneous disease severity. Although deep learning methods have shown promise for CT-based IPN prediction, most existing approaches underutilize severity-related information and provide limited spatial interpretability, restricting their robustness and clinical applicability. In this work, we propose a severity-consistent weakly supervised framework for CT-based IPN prediction with class activation mapping (CAM)-guided lesion localization. The framework leverages routinely available heterogeneous clinical annotations, including scan-level IPN labels and the computed tomography severity index (CTSI), without requiring pixel-level infection annotations. A segmentation-guided multi-task learning strategy is employed to jointly model IPN prediction and CTSI estimation using pancreas-centered representations, thereby stabilizing feature learning under weak supervision. To incorporate clinical severity knowledge, a ranking-based severity consistency regularization is introduced to encourage a monotonic relationship between disease severity and predicted IPN risk. In addition, a classification-to-localization weak supervision mechanism is developed, in which anatomically constrained CAMs are distilled into lesion representations and reintegrated into IPN prediction through lesion-aware pooling. Experimental results demonstrate that the proposed framework substantially improves IPN prediction performance, achieving an area under the receiver operating characteristic curve of 0.842 with a sensitivity of 0.796, while simultaneously providing visually plausible lesion localization. These findings suggest that the proposed approach offers a practical and interpretable solution for CT-based IPN risk assessment under limited annotation settings.