Mass detection from mammographic images using deep learning methods

Cancer refers to a serious group of diseases caused by the uncontrolled growth of cells and the metastasis process, in which these cells spread throughout the body and may lead to death. Early diagnosis of such diseases is highly important for effective treatment and prolonged patient survival. Today, breast cancer is mostly diagnosed by radiologists through the evaluation of mammography images. However, this process may vary due to its interpretative nature and is prone to human error. At this point, AI-supported decision systems can speed up the diagnostic process and improve reliability by increasing accuracy. This thesis investigates the mass detection performance of YOLOv8 and YOLOv7 architectures, which are based on the deep learning-based object detection algorithm YOLO (You Only Look Once), on mammography images. Due to their ability to detect objects in real time and with high precision, YOLO-based architectures are frequently used in medical image analysis. Three datasets were used in this study. The first dataset, containing 62 mammography images, was obtained from Dicle University Faculty of Medicine Hospital with ethics committee approval. These images were labeled by a radiologist using the LabelImg application. The second dataset consists of 162 images obtained by applying data augmentation to the original dataset. The third dataset includes 251 publicly available images from the Roboflow platform. The models were evaluated using performance metrics such as Mean Average Precision (mAP), Precision, Recall, and F1 Score. The results show that YOLOv8 performed more successfully and reliably than YOLOv7 under limited labeled data conditions. In all datasets, YOLOv8 achieved better outcomes in all metrics. In conclusion, the findings demonstrate that YOLOv8 is a strong candidate for medical diagnostic support systems due to its advanced architecture and experimental performance.