Detection and classification of lesions in mammogram images with deep learning

Breast cancer, one of the commonly occurring cancer types in human life, is more prevalent in women compared to men. Although the exact causes of this cancer are not fully understood, various factors such as dietary habits, menstrual status, childbirth, and the use of birth control pills are believed to contribute to its development. Early detection of breast cancer is crucial for effective treatment. There are various methods for detecting breast cancer, one of which involves Magnetic Resonance (MR) imaging techniques. In this study, three different pre-trained CNN models, namely Inception V2, Resnet101, and Resnet-50, were used in conjunction with the Faster R-CNN deep learning method for lesion detection and classification in Mammogram images. A total of 500 mammography screening images, consisting of cancerous and normal images, were obtained from "The Digital Database for Screening Mammography - DDSM," an open-source digital database. Out of these images, 400 were annotated by radiologists, while 100 were unmarked mammogram images of unrelated cases. The lesion regions in the annotated 400 cancerous images were labeled using the LabelImg annotation tool and stored as XML files. Subsequently, these labeled images were used as the training dataset. The training of each of the three CNN models used in lesion detection was performed with 300,000 steps and a batch size of 1, in conjunction with the Faster R-CNN deep learning method. As a result of these experiments, the Inception V2 model achieved an accuracy of 97.16%, the ResNet101 model achieved an accuracy of 94.74%, and the best-performing ResNet-50 model achieved an accuracy of 99.63%. These results can provide valuable support for early diagnosis by medical experts in this field. Furthermore, in the thesis, pre-trained CNN models were also used for benign and malignant tumor classification using the MIAS dataset. The training set consisted of 2016 benign and 4088 malignant images, while the validation set comprised 504 benign and 1024 malignant images. The training using the pre-trained models showed high success rates, as observed in the results of this study. Keywords: Breast cancer, Deep Learning, Faster-RCNN, ResNet, Inception