Scholarlink

<p>Deep learning for invasive lung adenocarcinoma subtyping remains vulnerable to real-world imaging perturbations. We present a margin consistency framework evaluating 203,226 patches from 143 whole-slide images across five adenocarcinoma subtypes from the BMIRDS-LUAD dataset. Our approach integrates attention-weighted aggregation with margin-aware training, achieving strong feature-logit space alignment with Kendall correlations of 0.88 during training and 0.64 during validation. To address feature over-clustering from contrastive regularization, we introduce Perturbation Fidelity scoring that applies structured perturbations through Bayesian-optimized parameters. Vision Transformer-Large achieves 95.20±4.65 percent accuracy, representing 40% error reduction from 92.00±5.36% baseline. ResNet101 with attention mechanism reaches 95.89±5.37% from 91.73±9.23% baseline, a 50% error reduction. All five subtypes exceed 0.99 area under receiver operating characteristic curves. Cross-institutional validation on WSSS4LUAD achieves 80.1% accuracy with ResNet50 with attention mechanism, demonstrating method robustness despite approximately 15–20% performance degradation from domain shift, indicating opportunities for future domain adaptation research.</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

Robust Histopathology Subtyping via Perturbation Fidelity in Deep Classifier

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Meghdad Sabouri Rad,
  • Junze Vincent Huang,
  • Mohammad Mehdi Hosseini,
  • Rakesh Choudhary,
  • Harmen Siezen,
  • Tamara Jamaspishvili,
  • Saverio J. Carello,
  • Ola El-Zammar,
  • Michel R. Nasr,
  • Bardia Rodd

摘要

Deep learning for invasive lung adenocarcinoma subtyping remains vulnerable to real-world imaging perturbations. We present a margin consistency framework evaluating 203,226 patches from 143 whole-slide images across five adenocarcinoma subtypes from the BMIRDS-LUAD dataset. Our approach integrates attention-weighted aggregation with margin-aware training, achieving strong feature-logit space alignment with Kendall correlations of 0.88 during training and 0.64 during validation. To address feature over-clustering from contrastive regularization, we introduce Perturbation Fidelity scoring that applies structured perturbations through Bayesian-optimized parameters. Vision Transformer-Large achieves 95.20±4.65 percent accuracy, representing 40% error reduction from 92.00±5.36% baseline. ResNet101 with attention mechanism reaches 95.89±5.37% from 91.73±9.23% baseline, a 50% error reduction. All five subtypes exceed 0.99 area under receiver operating characteristic curves. Cross-institutional validation on WSSS4LUAD achieves 80.1% accuracy with ResNet50 with attention mechanism, demonstrating method robustness despite approximately 15–20% performance degradation from domain shift, indicating opportunities for future domain adaptation research.