Cytokeratin Immunohistochemistry-Supervised Deep Learning for Detecting Breast Cancer Lymph Node Metastases and Evaluation of its Clinical Utility

Lymph node status is an indispensable examination for breast cancer therapy. To detect small, inconspicuous metastatic carcinomas, pathologists usually require immunohistochemical (IHC) staining for cytokeratin (CK). Here, we proposed an IHC-supervised algorithm to create virtual CK masks in lymph node hematoxyline and eosin (H&E) images and evaluated its clinical utility. We enrolled 194 patients with breast cancer surgery-related axillary lymph nodes containing variously sized metastases. The deep learning network, Unet++ with EfficientNet-B7 as the backbone, was trained with the ground truth extracted from consecutive or re-stained CK. At the pixel level, the model had high accuracy (0.98 on average) and decent recall (0.64 on average) and performed best in macrometastasis, followed by micrometastasis and isolated tumor cells (ITC). At the whole-slide image (WSI) level, all 25 slides with macro-metastases and most micro-metastatic (15/16) were classified correctly. For ITC, 17/19 patients were identified; however, certain benign cells were misrecognized in 18/19 negative patients. In clinical settings, artificial intelligence can help pathologists detect micrometastatic carcinoma and significantly decrease reading time. IHC-supervised deep learning is robust and efficient, providing substantial, high-quality ground truth. The virtual CK masks and augmented WSI system enhanced pathologists’ ability to search for tumors in the lymph nodes.