A First Step Towards an Algorithm for Breast Cancer Reoperation Prediction Using Machine Learning and Mammographic Images

Abstract: Cancer is the second leading cause of death worldwide and 30% of all cancer cases among women are breast cancer. A popular treatment is breast-conserving surgery, where only a part of the breast is surgically removed. Surgery is expensive and has a significant impact on the body, and on some women, a reoperation is needed. The aim of this thesis was to see if there is a possibility to predict whether a person will be in need of reoperation with the help of whole mammographic images and deep learning. \\ The data used in this thesis were collected from two different open sources: (1) The Chinese Mammography Database (CMMD) where 1052 benign images and 1090 malignant images were used. (2) The Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) where 182 benign images and 145 malignant images were used. With those images, both a simple convolutional neural network (CNN) and a transfer learning network using the pre-trained model MobileNet were trained to classify the images as benign or malignant. All the networks were evaluated using learning curves, confusion matrix, accuracy, sensitivity, specificity, AUC and a ROC-curve. The highest results obtained belonged to a transfer learning network that used the pre-trained model MobileNet and trained on the CMMD data set. It got an AUC value of 0.599.