Classification des images mammographiques selon leurs densités

Abstract

Breast cancer is a serious pathology in women, where early detection is of prime importance, as it considerably increases the chances of survival. Mammography is an essential technique in this process, but its interpretation is complex due to the variability of breast density. Currently, machine learning techniques, notably Deep Learning, are widely used for image classification. The main aim of this final-year project is to classify digital mammograms according to their density (B or C). The experiments in this project were carried out on the RSNA (The Radiological Society of North America) mammography database, consisting of 24350 mammograms. To reduce noise, some pre-processing techniques were applied, and classification was done based on transfer learning through pre- trained models such as VGG16, ResNet50, MobileNetV2 and EfficientNetB0. Among these models, EfficientNetB0 was elected as the best classification model because of these good results. This model achieved a recognition rate of 85.64%, a precision of 86%, a recall of 88% and an F1 score of 86%.