Predicting-Breast-Cancer-Malignancy-from-X-rays

Regarding the data, all 628 screening mammograms in this project have been classified as abnormal by 2 radiologists and thus require a biopsy. Radiologists are cautious during screenings, the consequences of having a false negative push them to send women for biopsy if there is the slightest doubt for them to have cancer. As explained, this results in a high number of false positive since the ‘cost’ is lower than having a false negative. The dataset used in this project comes from the OPTIMAM Medical Image Database, which collects NHS Breast Screening Programme (NHSB-SO) images in the UK. A deep learning approach is used to classify abnormal screenings as either malignant or benign cancer with a certain probability. Transfer Learning makes it possible to obtain high performances on small datasets. This project achieved a ROC of 80%, 86% sensitivity, and 77% NPV, which were reached with a pre-trained ResNet50v2, a state-of-the-art neural network optimized through fine-tuning hyperparameters and data pre-processing.