Breast cancer prediction from mammography images using deep learning techniques

Technological advancements have enabled the integration of intelligent machines into decision-making processes in various sectors, including healthcare. With their ability to perform analytical thinking and detect relationships faster than humans, these machines provide significant support to experts in making accurate decisions. The critical importance of early cancer diagnosis has accelerated artificial intelligence research in this field. Early cancer diagnosis is crucial for increasing patient survival rates, preventing organ loss, and minimizing the need for treatments such as radiotherapy and chemotherapy. The detection of small lesions in the breast is often challenging, leading to the disease progressing and being detected at advanced stages. Contrast-Enhanced Spectral Mammography (CESM) is an innovative imaging method designed to address the challenges associated with detection. By making lesions visible with a contrast agent injected into the breast cells, CESM helps to capture lesions that may be missed in digital mammography. It allows doctors to make more reliable diagnoses and detect the disease early. This thesis centers on diagnosing breast cancer through the analysis of Contrast-Enhanced Spectral Mammography (CESM) images. The study analyzed various deep learning models trained on two different datasets. The datasets used were the CDD-CESM dataset and the CESM@UCBM dataset. The CDD-CESM dataset contains a total of 2006 images, with 1003 low-energy images and 1003 subtracted images. The CESM@UCBM dataset contains 1138 images, with 569 low-energy and 569 subtracted images. In the training conducted within the scope of this study, data sets were approached from 20 different angles. The dataset was trained as 3-class (Normal, Benign, Malignant) and 2-class (Benign, Malignant), separately and together. Data augmentation methods were also applied to these datasets and the data was retrained. In this context, ResNet50, DenseNet121, EfficientNetB0, VGG16 and VGG19 architectures were applied, and as a result of five-fold cross-validation, an accuracy of 76.46%, sensitivity of 95.35% and F1-score of 84.81% were obtained with the DenseNet121 model. The findings suggest that the proposed approach could serve as a valuable tool in aiding the diagnosis of breast cancer. Keywords: Deep learning, Transfer learning, Breast cancer, Contrast-enhanced spectral mammography.