Scholarlink

Model brittleness across datasets is a key concern when deploying deep learning models in real-world medical settings. One approach is to fine-tune the model on subsequent datasets after training on the original dataset. However, this degrades model performance on the original dataset, a phenomenon known as catastrophic forgetting. We develop an approach to address catastrophic forgetting by combining elastic weight consolidation with a simple yet novel modulation of global batch normalization statistics under two scenarios: expanding the domain across 1) imaging systems and 2) hospital institutions. Focusing on the clinical use case of mammographic breast density detection, we show that our approach empirically outperforms several other state-of-the-art approaches and provides theoretical justification for the efficacy of batch normalization modulation, demonstrating the potential of our approach to deploying clinical deep learning models requiring domain expansion.

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

Addressing Catastrophic Forgetting by Modulating Global Batch Normalization Statistics for Medical Domain Expansion

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sharut Gupta,
  • Ken Chang,
  • Liangqiong Qu,
  • Aakanksha Rana,
  • Syed Rakin Ahmed,
  • Mehak Aggarwal,
  • Nishanth Arun,
  • Ashwin Vaswani,
  • Shruti Raghavan,
  • Vibha Agarwal,
  • Mishka Gidwani,
  • Katharina Hoebel,
  • Jay Patel,
  • Charles Lu,
  • Christopher P. Bridge,
  • Daniel L. Rubin,
  • Jayashree Kalpathy-Cramer,
  • Praveer Singh

摘要

Model brittleness across datasets is a key concern when deploying deep learning models in real-world medical settings. One approach is to fine-tune the model on subsequent datasets after training on the original dataset. However, this degrades model performance on the original dataset, a phenomenon known as catastrophic forgetting. We develop an approach to address catastrophic forgetting by combining elastic weight consolidation with a simple yet novel modulation of global batch normalization statistics under two scenarios: expanding the domain across 1) imaging systems and 2) hospital institutions. Focusing on the clinical use case of mammographic breast density detection, we show that our approach empirically outperforms several other state-of-the-art approaches and provides theoretical justification for the efficacy of batch normalization modulation, demonstrating the potential of our approach to deploying clinical deep learning models requiring domain expansion.