基于金字塔结构与注意力机制改进的YOLO算法在乳腺肿瘤超声图像检测中的应用研究

打开文本图片集

ABSTRACTTo improve the detectionaccuracy of benign and malignant breast tumors in ultrasound images，this study proposes adeep learning detection method based onanimproved YOLOv5s algorithm.A total of 659 breasttumor ultrasound imagesfromapubliclyavailable Kagledataset werecolected，withlesionsanotatedusing theLabelimg tool.Theimages were divided intoatrainingset（440images）andavalidationset（189images）ina7：3ratio，andtheremaining30imagesservedas an independent testset.Buildingupontheoriginal YOLOv5algorithm，this study introduced an Atrous Spatial Pyramid Pooling （ASPP）structure to enhance multi-scale feature extractionanda Coordinate Atention（CA）mechanism to improve lesion region focus.Three improved models were derived：YOLOv5s-AS（modifiedonlywithASPP），YOLOv5s-C（modifiedonly with CA mechanism），YOLOv5s-AS-C（optimized withboth ASPPand CA mechanisms）.After 20O roundsof iterative training，the mean average precision （mAP） of YOLOv5s-AS，YOLOv5s-C，and YOLOv5s-AS-C was 67.5%，77.5%，and 75.3% ， respectively，which was higher than that of the original YOLOv5s model（ 69.0% ）.Among them，YOLOv5s-C achieved the highest mAP.Testing onthebest-performing modelof YOLOv5s-Cdemonstratedsuperiorlesiondetectionaccuracy，whichachieved morepreciselesion localization fortheundetectedlesions intheoriginal model.Resultsof thisstudydemonstrated thatthe

YOLOv5smodel integrating CA mechanisms，significantly enhances breast tumor detection performance.Combining dep learning with manualfeature extractionachieves higherdetectionaccuracyandconfidence.Theproposed method canachieve eficientandauratedetectionandidentificationoflesionsinbreasttumorultrasoundimages，nabling moredirectclearand objective diagnosis，thereby enhancing diagnostic efficiency.

KEY WORDSUltrasound images;Artificial intelligence;Breast tumors;YOLO algorithm

乳腺癌是全球女性发病率最高的恶性肿瘤。（剩余7154字）