Whole slide histologic grading of breast cancer using convolutional neural networks.

Abstract

e13607 Background: The Nottingham histologic grade (NHG) is a strong prognostic factor in early-stage breast cancer. It consists of nuclear pleomorphism, tubular formation, and mitotic count. We recently developed an artificial intelligence (AI) based automatic grading system. Methods: In this study, we have retrospectively evaluated 179,651 hematoxylin and eosin-stained patches extracted from 402 digitized biopsies of 338 patients with confirmed invasive ductal carcinoma diagnosis. The data was collected at Acıbadem University Hospital between 2017 and 2021. The digitized slides were manually labeled by seven pathologists before being used to train the deep learning models (DL). The pre-trained (on ImageNet) DL architectures which are EfficientNet backbone U-Net, YOLOv5, DenseNet161, and a modified VGG-11 have been fine-tuned with this study’s dataset for the tubule segmentation, nuclei detection, mitosis classification, and the nuclear pleomorphism classification tasks, respectively. Data augmentation for boosting the accuracy was done. Semantic segmentation for tubular formation, object detection for detection of nuclei, and image classification for the mitotic count and nuclear pleomorphism was also performed. Results: The AI-based algorithms have obtained reproducible histologic scores with mean F1 scores, sensitivities, and specificities as presented in Table. Conclusions: The AI-based automatic grading system is accurate in the evaluation of the components of NHG. This system will speed up the pathology workflow in the clinic, provide decision support to pathologists and mitigate the sensitivity associated with the traditional grading process. [Table: see text]