YOLO as a Region Proposal Network for Diagnosing Breast Cancer

Cytological images of various types are increasingly being classified with the use of neural networks. But deep learning-based image classification systems are heavily reliant on manually sampled RoI (Region of Interest) patches. A lot of time and effort are required to extract RoI patches from whole slide images or larger images that are too complex to be processed by neural networks. A region proposal network (RPN) is an efficient way to automate the extraction of RoIs. In this study, we have proposed the use of the YOLOv3 network as an RPN to suggest RoIs in images from fine needle aspiration cytology of breast tissue. Patches from the suggested RoIs are fed into a Convolutional Neural Network (CNN) for the classification of benign and malignant lesions and ultimately, the diagnosis of Ductal Carcinoma in breast. The YOLO+CNN model yields a highly satisfactory classification accuracy of 95.73%, 100% specificity, 92.4% sensitivity and a precision score of 1.