SMDetector: Small mitotic detector in histopathology images using faster R-CNN with dilated convolutions in backbone model

Abstract

Breast cancer is one of the most common cancer types among women, and it is a deadly disease caused by the uncontrolled proliferation of cells. Pathologists face a challenging issue of mitotic cell identification and counting during manual detection and identification of cancer. Artificial intelligence can help the medical professional with early, quick, and accurate diagnosis of breast cancer. Consequently, the survival rate will be improved and mortality rate will be decreased. Different deep learning techniques are used in computational pathology for cancer diagnosis. In this study, the SMDetector is proposed which is a deep learning model for detecting small objects such as mitotic and non-mitotic nuclei. This model employs dilated layers in the backbone to prevent small objects from disappearing in the deep layers. The purpose of the dilated layers in this model is to reduce the size gap between the image and the objects it contains. Region proposal network is optimized to accurately identify small objects. The proposed model yielded overall average precision (AP) of 50.31% and average recall (AR) of 55.90% that outperforms the existing standard object detection models on ICPR 2014 (Mitos-Atypia-14) dataset. To best of our knowledge the proposed model is state-of-the-art model for precision and recall of mitotic object detection on ICPR 2014 (Mitos-Atypia-14) dataset. The proposed model has achieved average precision for mitotic nuclei 68.49%, average recall for mitotic nuclei 59.86% and f-measure 63.88%.