Comparing YOLOv8 and YOLOv9 Algorithm on Breast Cancer Detection Case
Ryan Marchi, S.S.F. Hau, Kristien Margi Suryaningrum, Rezki Yunanda · Procedia Computer Science · 2024
Breast cancer is a major global health concern, and early detection is crucial for positive patient outcomes. Mammography remains the primary screening tool, but radiologist shortages and heavy workloads create opportunities for computer-aided detection (CAD) systems. This study evaluates the performance of the state-of-the-art object detection models YOLOv8 and YOLOv9 on breast cancer detection and classification from mammography images in the CBIS-DDSM dataset. While YOLOv8 models have been explored for breast cancer detection, YOLOv9 with its novel architectural improvements like Programmable Gradient Information and the Generalized Efficient Layer Aggregation Network has not been previously benchmarked for this task. Our study uses a preprocessing of the CBIS-DDSM dataset which contain 1,514 images containing 850 benign and 768 malignant lesions. We have also applied data augmentation techniques such flipping horizontal and rotate 15 degree which increase the amount of dataset to 4498 images. In this study YOLOv8s, YOLOv8l, and YOLOv9c models were trained, and their performance compared. The YOLOv9c model achieved the highest mean average precision (mAP 50 ) of 0.677 and recall of 0.686, outperforming YOLOv8s and YOLOv8l. However, YOLOv8l attained the highest precision of 0.788. These results demonstrate the promising potential of advanced deep learning algorithms for breast cancer computer-aided diagnosis. Further research should focus on expanding training data, robust preprocessing, and integrating these models into clinical workflows for validating their practical utility in improving breast cancer diagnosis and patient care.