AI-AUGMENTED PATHOLOGY: THE EXPERIENCE OF TRANSFER LEARNING AND INTRA-DOMAIN DATA DIVERSITY IN BREAST CANCER METASTASIS DETECTION

Manuel Cossio^1*Nina Wiedemann²Esther Sanfeliu Torres³Esther Barnadas Sole³Laura Igual¹

• ¹University of Barcelona, Barcelona, Spain
• ²ETH Zürich, Zurich, Zürich, Switzerland
• ³Hospital Clinic of Barcelona, Barcelona, Catalonia, Spain

Background: Metastatic detection in sentinel lymph nodes remains a crucial prognostic factor in breast cancermanagement, with accurate and timely diagnosis directly impacting treatment decisions. While traditionalhistopathological assessment relies on microscopic examination of stained tissues, the digitization of slidesas Whole Slide Images (WSI) has enabled the development of computer-aided diagnostic systems. Theseautomated approaches offer potential improvements in detection consistency and efficiency compared toconventional methods.Results: This study leverages transfer learning on hematoxylin and eosin (H&E) whole-slide images (WSIs)to achieve computationally efficient metastasis detection without compromising accuracy. We propose anapproach for generating segmentation masks by transferring spatial annotations from immunohistochemistry(IHC) WSIs to corresponding H&E slides. Using these masks, four distinct datasets were constructed tofine-tune a pretrained ResNet50 model across eight different configurations, incorporating varied datasetcombinations and data augmentation techniques. To enhance interpretability, we developed a visualization toolthat employs color-coded probability maps to highlight tumor regions alongside their prediction confidence.Our experiments demonstrated that integrating an external dataset (Camelyon16) during training significantlyimproved model performance, surpassing the benefits of data augmentation alone. The optimal model,trained on both external and local data, achieved an accuracy and F1-score of 0.98, outperforming existingstate-of-the-art methods.Conclusions: This study demonstrates that transfer learning architectures, when enhanced with multi-sourcedata integration and interpretability frameworks, can significantly improve metastatic detection in wholeslide imaging. Our methodology achieves diagnostic performance comparable to gold-standard techniqueswhile dramatically accelerating analytical workflows. The synergistic combination of external dataset incor-poration and probabilistic visualization outputs provides a clinically actionable solution that maintains bothcomputational efficiency and pathological interpretability.