Improved YOLO: application of feature enhancement and multi-scale fusion in breast tumor ultrasound image detection

Objective: Optimize and upgrade the detection method of breast tumor ultrasound images using an improved YOLO-based algorithm model. Methods: A total of 659 breast tumor images were selected from the Kaggle database as the initial dataset. The detection targets in these images were pre-annotated using the image annotation tool, Labelimg. Randomly selecting 629 images, they were divided into training and validation sets in a ratio of 7:3. The remaining 30 images were used as the test set. To improve upon the original YOLO algorithm, we introduced SE attention mechanism and BiFPN for structural refinement, resulting in a modified model named YOLOv5-BiFPN-SE. Both the training and validation sets were used to train both the original YOLO model and YOLOv5-BiFPN-SE model through 150 iterations. The best weight file obtained from this process was then utilized for final evaluation on the test set. Results After 150 rounds of iterative training, the average accuracy rates for detecting breast tumors in ultrasound images were found to be 72.1% and 78.1% for the two models respectively. The improved model exhibited a significant enhancement in its ability to detect small targets within the images compared to the original model. Conclusion: Compared to the original YOLO model, the improved YOLO model enhances the accuracy and sensitivity in recognizing small targets in breast tumor ultrasound images, thereby contributing to improving diagnostic efficiency of breast tumors in clinical settings.