Scholarlink

Multi-channel approaches in medical imaging AI prioritize accuracy, but their impact on algorithmic robustness across clinically critical subgroups remains underexplored. We present a systematic comparison between single-channel (post-contrast) and multi-channel (pre+post+subtraction) strategies using the 306 test cases from the MAMA-MIA Challenge dataset, which spans four institutions. Despite similar overall performance (Dice: 0.743 vs 0.739, p = 0.67), single-channel models exhibited superior robustness for molecular subtypes that require targeted therapies. HR-/HER2+ cases improved across all five robustness metrics (Cohen’s d = 0.320), while HR+/HER2+ improved in 4/5 metrics. For patients receiving anti-HER2 therapies costing over $100,000/year (n = 93), single-channel segmentation achieved reliable detection (Dice  \(\ge \)  0.5) in 82.8% of cases versus 80.6% with multi-channel. Bootstrap validation (n = 10,000) confirmed these advantages, with 95% confidence intervals supporting consistency improvements. These findings challenge the assumption that greater input complexity improves clinical reliability, demonstrating that input simplification can enhance robustness and guide equitable AI deployment in precision oncology.

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

Input Simplification Impact on Robustness for Targeted Therapy Subtypes in Breast MRI Segmentation AI

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Alba Bernal Rodriguez,
  • Beatriz Remeserio,
  • Justin Engelmann,
  • Lucas Gago,
  • Jacinto Velasco Rebolledo

摘要

Multi-channel approaches in medical imaging AI prioritize accuracy, but their impact on algorithmic robustness across clinically critical subgroups remains underexplored. We present a systematic comparison between single-channel (post-contrast) and multi-channel (pre+post+subtraction) strategies using the 306 test cases from the MAMA-MIA Challenge dataset, which spans four institutions. Despite similar overall performance (Dice: 0.743 vs 0.739, p = 0.67), single-channel models exhibited superior robustness for molecular subtypes that require targeted therapies. HR-/HER2+ cases improved across all five robustness metrics (Cohen’s d = 0.320), while HR+/HER2+ improved in 4/5 metrics. For patients receiving anti-HER2 therapies costing over $100,000/year (n = 93), single-channel segmentation achieved reliable detection (Dice  \(\ge \)  0.5) in 82.8% of cases versus 80.6% with multi-channel. Bootstrap validation (n = 10,000) confirmed these advantages, with 95% confidence intervals supporting consistency improvements. These findings challenge the assumption that greater input complexity improves clinical reliability, demonstrating that input simplification can enhance robustness and guide equitable AI deployment in precision oncology.