Performance evaluation of deep learning based YOLOv5 and YOLOv8 models for real time breast cancer detection in mammographic images.

Authors

Abstract

Breast cancer detection remains a significant challenge in medical diagnostics. Traditional diagnostic methods are time-consuming, unable to detect complex patterns in medical images, and achieve only moderate accuracy with high false-positive rates. The study aims to identify the most effective and efficient algorithm for clinical application that reduces the rates of false positives and false negatives Various image segmentation methods are available, including U-Net (Ronneberger et al. in International Conference on Medical image computing and computer-assisted intervention, Springer International Publishing, Cham, 2015), Mask R-CNN (He et al. in Proceedings of the IEEE international conference on computer vision, 2017), fully convolutional networks (FCNs) (Long et al. in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015), and SegNet (Badrinarayanan et al. in IEEE Trans Pattern Anal Mach Intell, 39:2481-2495, 2017). However, some of these have complex architectures, require extensive training time, and consume significant computational resources. Beyond these methods, other well-cited studies include CNN & context aggregation, transformer-based segmentation, promptable models, and medical-focused architectures. However, all of these still struggle with issues like long-range dependencies caused by the locality of convolutions, architecture design, high computational cost, require large amounts of pre-training data, not always being optimal for domain-specific tasks, high compute and memory demands, and the need for fine-tuning for specialised high-accuracy applications that are frequently trained on relatively small datasets. Therefore, this study employs YOLO, which is a DL-based one-step object detection algorithm that greatly enhances speed, accuracy, and recognition across various categories of image processing and video processing. It also lies in its ability to perform real-time object detection and its time-sensitive applications. This study assesses the effectiveness of two cutting-edge deep learning based object detection algorithms, YOLOv5 and YOLOv8, for early breast cancer screening and detection. A total of 2,620 digitised mammography images from film, including normal, benign, and malignant cases with verified pathology, are included in the original Digital Database for Screening Mammography (DDSM) dataset. Improving patient outcomes and early detection rates depends on the efficacy of these algorithms.YOLOv5 and YOLOv8 are used in this study to detect breast cancer lesions, and their performance metrics are compared. The results for YOLOv5 show an F1 score of 0.97, a precision-recall score of 0.97, and confidence scores for precision and recall of 0.67 and 0.98, respectively. The results for YOLOv5 show an F1 score of 0.97, a precision-recall score of 0.97, and confidence scores of 0.67 and 0.98 for precision and recall, respectively. The mAP of YOLOv8 is 0.99. Therefore, YOLOv8 outperforms YOLOv5 across all assessed parameters, indicating that it is more suitable for clinical use in automated breast cancer screening and detection. With YOLOv8's improved recall and precision, there may be a greater likelihood of accurate, early detection of breast cancer, thereby reducing false positives and negatives. Future research will refine these models for real-time use to improve comprehensive cancer screening programs and investigate their integration with other diagnostic techniques.

Tags

Topics