Automated Diagnosis of Breast Cancer with RoI Detection Using YOLO and Heuristics

Ananya Bal, Meenakshi Das, Shashank Mouli Satapathy, Madhusmita Jena, Subha Kanta Das

Abstract: Breast Cancer (specifically Ductal Carcinoma) is widely diagnosed by Fine Needle Aspiration Cytology (FNAC). Deep Learning techniques like Convolutional Neural Networks (CNNs) can automatically diagnose this condition by processing images captured from FNAC. However, CNNs are trained on manually sampled RoI (Region of Interest) patches or hand-crafted features. Using a Region Proposal Network (RPN) can automate RoI detection and save time and effort. In this study, we have proposed the use of the YOLOv3 network as an RPN and supplemented it with image-based heuristics for RoI patch detection and extraction from cytology images. The extracted patches were used to train 3 CNNs - VGG16, ResNet-50 and Inception-v3 for classification. YOLOv3 identified 164 RoIs in 26 out of 27 images and we achieved 96.6%, 98.8% and 98.9% classification accuracies with VGG16, ResNet-50 and Inception-v3 respectively.