Consecutive Cytokeratin Immunochemistry-Supervised Algorithm for Predicting Tumor Areas in Ki67 Breast Cancer Images

Abstract

Automatic Ki67 index (KI) assessment has become popular in breast cancer research; however, the results are easily influenced by non-tumor cells. This can be addressed by using neural networks to predict tumor areas. Compared to human annotation, cytokeratin immunostaining can more accurately highlight epithelial regions and provide reliable ground truth. We built an immunohistochemistry (IHC)-supervised neural network using the ground truth extracted from consecutive cytokeratin-stained slides, which could predict the tumor area in Ki67 images of breast cancer. The effect of masks on KI quantification was evaluated in 20 patients with breast carcinoma. Set A (three cases) was used to measure the similarity of adjacent whole-slide images (WSIs). A UNet++ (with an EfficientNet-b7 backbone) model was built using Set B (67 cases) for tumor area prediction. The KI in Set C (20 cases) was quantified with and without the application of tumor-area masks, and the KI difference was computed. The mean intersection over union of the epithelial masks extracted from adjacent cytokeratin sections was 0.72 (0.68–0.76). After training and validating in 49 cases, the intersection over union in the test set was 0.44–0.73. At the tile image-level, KI difference was −42.5–41.7%. Images with the highest difference usually contained numerous lymphocytes or vessels, and the masks prevented disguised cells from being counted. At the WSI-level, the hotspot location changed in 18/20 cases, but hotspot KI changed insignificantly (−1.0% on average). The global KI changed less (0.9% on average). Thus, consecutive IHC staining provides substantial, precise, and reliable ground truths that trained the algorithm efficiently. This IHC-supervised training workflow can be applied to other targets by replacing IHC antibodies. Furthermore, the virtual tumor areas improved Ki67 counting by excluding the non-tumor areas at the tile image-level, although the effect on the WSI-level was insignificant.