Scholarlink

<p>Fine-grained medical image classification is often challenged by scarce annotations, subtle inter-class differences, and strong dependence on anatomical context. Conventional CutMix-style augmentation may be less suitable for such tasks because random region replacement can violate anatomical plausibility and weaken the correspondence between pathology-related regions and diagnostic labels. Using vesicoureteral reflux (VUR) grading on voiding cystourethrography (VCUG) as a representative setting, we propose <b>MGMix</b>, a mask-guided anatomy-aware region mixing augmentation method. MGMix uses reflux-related masks to localize clinically relevant regions and performs aligned region mixing with task-specific hard-label transfer, producing augmented samples that better preserve anatomical structure and lesion semantics. Experiments on the curated VCUG dataset show that MGMix consistently improves left- and right-sided VUR grading across the two evaluated backbone architectures compared with the baseline and competing augmentation methods. These results demonstrate the value of anatomy-aware region mixing for data-scarce and structure-sensitive medical image classification tasks when reliable lesion or anatomical masks are available.</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

Mask-guided anatomy-aware region mixing for fine-grained medical image classification: VUR grading on VCUG

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Jiayuan Wang,
  • Ziyao Meng,
  • Shengwei Tian,
  • Yuyin Ma,
  • Zheyuan Wang

摘要

Fine-grained medical image classification is often challenged by scarce annotations, subtle inter-class differences, and strong dependence on anatomical context. Conventional CutMix-style augmentation may be less suitable for such tasks because random region replacement can violate anatomical plausibility and weaken the correspondence between pathology-related regions and diagnostic labels. Using vesicoureteral reflux (VUR) grading on voiding cystourethrography (VCUG) as a representative setting, we propose MGMix, a mask-guided anatomy-aware region mixing augmentation method. MGMix uses reflux-related masks to localize clinically relevant regions and performs aligned region mixing with task-specific hard-label transfer, producing augmented samples that better preserve anatomical structure and lesion semantics. Experiments on the curated VCUG dataset show that MGMix consistently improves left- and right-sided VUR grading across the two evaluated backbone architectures compared with the baseline and competing augmentation methods. These results demonstrate the value of anatomy-aware region mixing for data-scarce and structure-sensitive medical image classification tasks when reliable lesion or anatomical masks are available.