Publication

Abstract : World Health Organization (WHO) has reported that breast cancer is the most often found cancer in women and it is adversary affecting millions of women all around the world. Early detection and real-time screening can immensely assist the patient. Mitotic nuclei detection in breast histopathology images plays a critical function to evaluate the aggressiveness of the cancer malignancy. Cancer is identified by pathologists by analyzing histopathology images of tissues and determines numerous biomarkers. Since there is only minute variation among mitotic and not mitotic cells, this procedure is tedious, time-consuming, and instinctive. Various image processing techniques and deep learning models had been proposed to automate the procedure of detecting mitotic cells from the histopathology images. Traditional techniques commonly perform nuclei segmentation followed by classification which calls for immoderate computational resources. These models also lack expected accuracy due to the shortage of proper balanced datasets and errors during image staining. In this paper, we define the challenges as an object detection task, wherein the mitotic nuclei are directly predicted without nuclei segmentation in a single step using YOLOv4, which is a fast-operating object detection model. The model was trained with 506 mitosis instances from the openly available MITOS-ATYPIA-14 grand challenge dataset that comprises hematoxylin and eosin (H&E) stained breast histopathology images annotated by experienced pathologists. The outcome suggests that the YOLOv4 model with RGB images as input offers an F-measure of 0.73 and can be used as a dependable and much less computationally exhaustive approach among the prevailing ones.