YOLOv12-based Detection for Early Breast Cancer Screening with a Portable Ultrasound System

Authors:

Abstract
This study explores the application of deep learning for breast lesion detection in ultrasound images using the YOLOv12 object detection model. Leveraging a compact wireless ultrasound probe and an Android-based inference pipeline, the system was developed to enable portable, AI-assisted screening in resource-limited settings. The model was trained and evaluated on an annotated ultrasound dataset and compared against the previous YOLOv11 version. YOLOv12 achieved a mean average precision (mAP) of 90.6% and an F1 score of 88.65%, outperforming YOLOv11 in both accuracy and inference speed. Processing time was also reduced, with YOLOv12 achieving image detection times between 1.34 to 4.42 seconds, compared to YOLOv11’s slower range. These results confirm YOLOv12’s suitability for real-time deployment on mobile platforms. Visual analyses across several test images show that YOLOv12 offers more consistent detection across varying lesion sizes and positions. The system's lightweight design, combined with its robust performance, makes it a promising tool for expanding diagnostic access in rural and underserved regions. Future work will focus on multiclass lesion classification, expanded datasets, and clinical usability studies to further validate its application in real-world healthcare environments.

This publication has 19 references indexed in Scilit: