YOLOv8-Based Automatic Lesion Detection in Breast Ultrasound Images: A Reproducible Evaluation on a Small and Imbalanced Dataset

Abstract

Early detection is among the strongest determinants of survival in breast cancer, and ultrasound imaging is a routine component of the diagnostic pathway, particularly for patients with dense breast tissue. This study applies YOLOv8n, a lightweight single-stage object detector, to the task of localising lesions in breast ultrasound images and separating benign from malignant findings. Pixel-level lesion masks from the 780 images of the publicly available Breast Ultrasound Images (BUSI) dataset were converted into axis-aligned bounding boxes by external contour analysis, while lesion-free images were retained as unlabelled negative examples. Images were partitioned class-wise into training, validation and test subsets in a 70/15/15 ratio, and class-specific offline augmentation was applied to the training subset only, which brought the numbers of benign and malignant training instances to approximately the same level (614 versus 592). The model was trained for 50 epochs at an input resolution of 640×640. On the test subset, which was used neither for training nor for model selection, the mean average precision at an intersection-over-union threshold of 0.50 (mAP@50) was 0.737 and mAP@50–95 was 0.506. Class-wise mAP@50 was 0.817 for benign and 0.656 for malignant lesions. The findings indicate that a lightweight single-stage detector can reach a reasonable localisation performance on a small ultrasound dataset, but that performance on the malignant class — the clinically decisive one — is not yet adequate for decision-support use. The contribution of this work is not a claim of high accuracy but a reproducible account of every step from data preparation to evaluation, together with an explicit statement of the limitations of the approach.