Scholarlink

This study proposes an Ensemble Transformer-based Multiple Instance Learning (MIL) framework for predicting Neoadjuvant Chemotherapy (NAC) response in breast cancer using pre-treatment biopsy Whole Slide Images (WSIs). The model was developed on multi-center data from 128 patients (86 training, 42 internal validation) and externally validated on 22 microscope images. Utilizing ResNet50 for feature extraction and a multi-scale attention Transformer for encoding pathological information, the framework incorporated a two-stage classification strategy, domain adaptation, and ensemble learning. The model achieved 79.3% WSI-level accuracy in internal validation, with AUCs of 0.82 for pathological complete response (pCR) and 0.77 for non-response categories. External validation yielded AUCs of 0.70 for pCR and 0.67 for non-response, outperforming traditional models like GoogleNet, ResNet34, and SqueezeNet across all metrics. The proposed method demonstrates high accuracy and interpretability while relying solely on WSIs, making it suitable for resource-limited clinical settings and providing fine-grained identification of pCR and non-responsive patients for personalized treatment.

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

Ensemble Transformer-Based Multiple Instance Learning for Predicting Neoadjuvant Chemotherapy Response from Breast Cancer Biopsy Whole-Slide Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kaixin Du,
  • Kaining Ye,
  • Zhenshui Wu,
  • Bing Chen,
  • Junqiang Hong,
  • Zhongping Zhang,
  • Yongliang Dai,
  • Xuehong Liao

摘要

This study proposes an Ensemble Transformer-based Multiple Instance Learning (MIL) framework for predicting Neoadjuvant Chemotherapy (NAC) response in breast cancer using pre-treatment biopsy Whole Slide Images (WSIs). The model was developed on multi-center data from 128 patients (86 training, 42 internal validation) and externally validated on 22 microscope images. Utilizing ResNet50 for feature extraction and a multi-scale attention Transformer for encoding pathological information, the framework incorporated a two-stage classification strategy, domain adaptation, and ensemble learning. The model achieved 79.3% WSI-level accuracy in internal validation, with AUCs of 0.82 for pathological complete response (pCR) and 0.77 for non-response categories. External validation yielded AUCs of 0.70 for pCR and 0.67 for non-response, outperforming traditional models like GoogleNet, ResNet34, and SqueezeNet across all metrics. The proposed method demonstrates high accuracy and interpretability while relying solely on WSIs, making it suitable for resource-limited clinical settings and providing fine-grained identification of pCR and non-responsive patients for personalized treatment.