HER2-positive breast cancer accounts for about 20% of breast cancer cases and is generally associated with more aggressive behavior and a higher risk of distant metastasis, while pathologic complete response (pCR) after neoadjuvant chemotherapy is strongly linked to better survival outcomes. As a result, accurately predicting treatment response before therapy remains an important clinical challenge. The structural and spatial heterogeneity of the tumor microenvironment is increasingly recognized as a key determinant of response, and routine H&E pathology slides naturally preserve spatial information about tumor, stroma, and immune infiltration, making them a valuable source for prediction. Although immunohistochemical biomarkers are commonly used in clinical practice, these approaches are labor-intensive, time-consuming, and difficult to scale. More recent deep-learning methods based on whole-slide images have improved automation, but many still treat slides as collections of independent tiles, limiting their ability to explicitly model tissue-level spatial organization and reducing interpretability. "Graph-based methods have started to address spatial context, yet many still lack clear tissue-compartment semantics and rarely integrate spatial structural features, deep semantic features, and clinical variables within a unified framework." said the author Wensheng Cui, a researcher at Hangzhou Dianzi University, "This makes biologically grounded and more interpretable tissue-specific modeling an important direction for predicting treatment response in HER2-positive breast cancer."
This study proposed a hierarchical tissue-specific modeling framework to predict pathologic complete response to neoadjuvant chemotherapy in HER2-positive breast cancer from routine H&E whole-slide images. The workflow first divided each slide into 5 biologically meaningful tissue compartments, including tumor, stroma, stromal tumor-infiltrating lymphocytes (sTILs), intratumoral tumor-infiltrating lymphocytes (iTILs), and overall TILs. For each compartment, a tissue graph was then constructed using clustered representative tiles as nodes and spatial proximity as edges, from which interpretable spatial structural features were extracted through social network analysis. In parallel, deep semantic information was obtained using a pretrained weakly supervised multiple-instance learning model to generate tissue-specific deep-learning prediction scores, which were then integrated with clinical variables to build compartment-specific predictive models. The framework was trained on the Yale Response cohort and externally validated on the independent IMPRESS HER2+ dataset to systematically assess the contribution of different tissue compartments and feature combinations to treatment response prediction.
This hierarchical tissue-specific framework showed strong predictive performance for neoadjuvant chemotherapy response in HER2-positive breast cancer, with clear differences across tissue compartments. Among them, the stromal compartment achieved the best performance, reaching an AUC of 0.907 in the external validation cohort, outperforming previously reported methods as well as models based only on clinical variables, deep-learning scores, or simple tissue-count features. This suggests that the stroma, beyond the tumor itself, contains highly informative signals related to treatment response. Overall, integrating spatial structural features, deep semantic features, and clinical variables generally produced more stable and accurate predictions than any single feature source alone. Notably, the spatial graph features derived from social network analysis showed strong standalone predictive value across multiple compartments, and in the stromal compartment they even outperformed both deep-learning scores and clinical variables, highlighting spatial tissue organization itself as an important indicator of response. Further feature analysis also revealed compartment-specific patterns: the tumor compartment relied more heavily on deep semantic information, whereas stromal and immune-related compartments benefited more from spatial structural features. Together, these results show that tissue-specific hierarchical modeling can capture treatment-related heterogeneity in the tumor microenvironment more effectively than treating the whole slide as a single undifferentiated image.
This work presents a more biologically grounded and interpretable way to analyze pathology images by moving beyond black-box whole-slide prediction and instead modeling tumor, stromal, and immune-related compartments separately, then integrating spatial structure, deep semantic information, and clinical variables to predict response to neoadjuvant chemotherapy in HER2-positive breast cancer. Importantly, the findings show that the most informative signals are not confined to the tumor itself, as stromal and immune-related compartments also carry substantial treatment-response information, with the particularly strong performance of the stromal compartment further underscoring the importance of spatial organization within the tumor microenvironment. The value of this study lies not only in improving predictive accuracy, but also in offering a more explainable framework for using routine H&E slides in clinical decision support, suggesting that tissue-specific hierarchical modeling could become an important bridge between digital pathology and treatment planning. "At the same time, the framework is still based on relatively limited public cohorts and models spatial organization mainly at the tissue level, so further validation on larger multicenter datasets and integration with finer-grained cellular information will be important for making this approach more robust, more generalizable, and closer to real clinical translation." said Wensheng Cui.
Authors of the paper include Wensheng Cui, Tao Tan, Ming Fan, and Lihua Li.
This work is supported in part by the National Natural Science Foundation of China under grants W2411054, 62271178, and U21A20521 and by the Zhejiang Provincial Natural Science Foundation of China (LR23F010002).
The paper, "Hierarchical Tissue-Specific Modeling of Pathology Images Predicts Response in HER2+ Breast Cancer" was published in the journal Cyborg and Bionic Systems on Apr 22, 2026, at DOI: 10.34133/cbsystems.0554.
