Lightweight Detection Algorithm for Breast-Mass Features in Ultrasound Images
Real-time analysis of ultrasound videos using embedded terminals enables the rapid detection of breast masses and plays a crucial role in early breast cancer screening and diagnosis. However, as a paired organ with a large area, the breast consists of interwoven fatty layers, mammary ducts, glandula...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10979300/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Real-time analysis of ultrasound videos using embedded terminals enables the rapid detection of breast masses and plays a crucial role in early breast cancer screening and diagnosis. However, as a paired organ with a large area, the breast consists of interwoven fatty layers, mammary ducts, glandular tissues, and masses, leading to a high false-positive rate in target detection. Unlike standard imaging modalities such as computed tomography, X-ray, and magnetic resonance imaging, ultrasound detection heavily depends on the sonographer’s techniques and expertise, which can lower detection accuracy, particularly in identifying small masses. This paper proposes a novel architecture, LEW-YOLO, to address the real-time detection demands of embedded terminal devices for breast ultrasound. First, we introduce an efficient multiscale convolutional (EMSC) module to improve feature extraction from complex backgrounds and enhance multiscale representation. Unlike traditional methods, EMSC employs multiple convolution branches with varying kernel sizes, enabling more effective multiscale feature capture. Second, the detection head of YOLOv8 is replaced with a lightweight shared detail-enhanced convolutional detection head (LSDECD) to improve the model’s ability to detect small masses. Finally, a weighted intersection-over-union (WIoU) loss function is integrated to better capture the complex boundaries of malignant masses. Experiments on the public BUSI dataset demonstrate that LEW-YOLO achieves an mAP50 of 91.6%, surpassing the YOLOv8n baseline (89.0%) by 2.6%. On the BUET dataset, LEW-YOLO attains an mAP50 of 83.3%, outperforming YOLOv8 (72.3%) by 11.0%. Moreover, the parameter count and GFLOPs are reduced by 26.6% and 22.2%, respectively. Thus, the proposed model effectively balances detection performance and lightweight design, making it well-suited for real-time applications on resource-constrained computing devices. |
|---|---|
| ISSN: | 2169-3536 |