Detection, localization, and staging of breast cancer lymph node metastasis in digital pathology whole slide images using selective neighborhood attention-based deep learning.

Accurate detection, localization, and staging of breast cancer lymph node metastases are critical for guiding treatment decisions and predicting patient outcomes. This study presents a selective neighborhood attention-based deep learning framework that combines nuclei-level features with high-level tissue embeddings to detect, annotate and stage breast cancer metastases in whole-slide images (WSIs) of lymph node biopsy specimens precisely. The proposed framework leverages a dual-path feature extractor, incorporating both nuclei segmentation/classification outputs and transformer-based tissue features, alongside a dynamic attention mechanism that selects and emphasizes neighboring patches based on similarity to the target patch. Experimental results on the CAMELYON16 test set demonstrate high performance in patch-level tumor detection, with sensitivity of 96.2 ± 1.5%, precision of 95.3 ± 2.4%, and an F1-score of 95.7 ± 3.1%. The model achieves accurate tumor boundary delineation, evidenced by a Dice score of 90.5 ± 2.0% and a Jaccard index of 82.6 ± 0.8%, along with a lesion-level free-response receiver operating characteristic (FROC) score of 84.6 ± 2.8%. Additionally, the slide-level classification achieves an area under the receiver operating characteristic (ROC) curve (AUC) of 0.96 ± 0.01, highlighting the system's strong diagnostic capability. Out-of-distribution evaluation on the CAMELYON17 dataset confirms the framework's generalizability, yielding an F1-score of 87.0 ± 1.8% at the patch level and an AUC of 0.88 ± 0.03 at the slide level. Furthermore, the proposed model achieves a kappa score of 0.94 ± 0.02 for automated pN-staging at the patient level, indicating near-expert concordance in detecting and classifying the extent of nodal metastasis. Ablation analyses underscore the importance of incorporating nuclei-based features and selective neighborhood attention, with noticeable performance degradation observed when either element is removed. By integrating cellular-level insights with tissue-level contextual information, the proposed framework replicates key aspects of human pathological assessment effectively and shows promise as a decision-support tool in the era of digital pathology.