Breast Cancer Lesion Detection in Mammograms Using YOLO v8
Maryam Ahmed
Co-Presenters: Individual Presentation
College: The Dorothy and George Hennings College of Science, Mathematics and Technology
Major: Computer Science
Faculty Research Mentor: Kuan Huang
Abstract:
Breast cancer is among the leading causes of mortality in women and brings to light the need for methods of early and accurate detection. Traditional mammogram analysis is plagued with significant challenges, including the variability in the appearance of lesions and dense breast tissues, and the fact that they are all manual diagnoses that take inordinate time. In this study, we explore the YOLO (You Only Look Once) object detection framework to carry out breast cancer lesion detection in the parameters of mammograms.This research develops a solid breast cancer detection pipeline using the Emory BrEast Imaging Dataset (EMBED). The methodology includes preprocessing mammogram images, including converting DICOM files to PNG format, drawing annotations around lesions in red rectangles, and training YOLOv8 to identify those regions of interest. Evaluation of the performance of the models is undertaken on the key metrics of precision, recall, F1-score, and mean Average Precision (mAP). At the same time, improvement in detection accuracy and efficiency of the analyses is delivered through comparative performance with YOLOv11.The present study focuses on reaping the advantages of artificial intelligence in clinical imaging with the impact of optimizing breast cancer screening and making it a contribution toward furthering life-saving diagnostic automated solutions. Our results indicate that YOLOv8 identifies cancerous lesions with high reliability, as evidenced by clearly delineated annotated images. Future works include laying refinements on data preprocessing methods, extended model evaluation, and further adaptation by other capabilities within deep learning, such as attention mechanisms for further accuracy enhancement. We will also analyze the EMBED dataset in the future, including the impact of rheumatoid arthritis on mammogram positioning, as well as fairness across racial and ethnic groups. Additionally, we will examine different mammogram types, such as 2D, 3D, and C-view, to enhance AI model performance.