Scholarlink

<p>Cross-magnification variability poses a critical challenge in deploying AI models for digital pathology, as classifiers trained at a single resolution often fail to generalize across heterogeneous imaging conditions. This study systematically investigates the effects of magnification domain shifts on breast histopathology classification by combining handcrafted machine-learning features, modern deep architectures, and progressive domain adaptation strategies. Although handcrafted models performed well in the domain, they degraded sharply under cross-magnification transfer, and shallow alignment methods such as CORAL and MMD yielded only marginal improvements. In contrast, deep transformer and convolutional backbones demonstrated greater robustness through hierarchical, scale-tolerant feature learning. Adversarial domain adaptation with DANN yielded the most consistent cross-magnification generalization (AUC = 0.975) and preserved malignant recall, a critical clinical requirement. Gradient reversal strength experiments revealed that moderate <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\lambda\)</EquationSource> <EquationSource Format="MATHML"><math> <mi>λ</mi> </math></EquationSource> </InlineEquation> values (0.5–0.8) yield an optimal balance between domain invariance and diagnostic discrimination. Grad-CAM and t-SNE confirmed that adapted models focused on biologically meaningful patterns and maintained coherent feature geometry across magnifications. Overall, this work highlights deep adversarial adaptation as an effective and interpretable solution for magnification-agnostic, clinically transferable computational pathology.</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

Generalize Beyond the Lens: Deep Magnification Adaptation in Breast Cancer Histopathology

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月4日

查看全文
  • K. A. Shahriar,
  • Maruf Ahmed,
  • Sakib A. Hossain,
  • Muhammad E. H. Chowdhury,
  • Enamul H. Bhuiyan

摘要

Cross-magnification variability poses a critical challenge in deploying AI models for digital pathology, as classifiers trained at a single resolution often fail to generalize across heterogeneous imaging conditions. This study systematically investigates the effects of magnification domain shifts on breast histopathology classification by combining handcrafted machine-learning features, modern deep architectures, and progressive domain adaptation strategies. Although handcrafted models performed well in the domain, they degraded sharply under cross-magnification transfer, and shallow alignment methods such as CORAL and MMD yielded only marginal improvements. In contrast, deep transformer and convolutional backbones demonstrated greater robustness through hierarchical, scale-tolerant feature learning. Adversarial domain adaptation with DANN yielded the most consistent cross-magnification generalization (AUC = 0.975) and preserved malignant recall, a critical clinical requirement. Gradient reversal strength experiments revealed that moderate \(\lambda\) λ values (0.5–0.8) yield an optimal balance between domain invariance and diagnostic discrimination. Grad-CAM and t-SNE confirmed that adapted models focused on biologically meaningful patterns and maintained coherent feature geometry across magnifications. Overall, this work highlights deep adversarial adaptation as an effective and interpretable solution for magnification-agnostic, clinically transferable computational pathology.