Vanilla-YOLO: A Lightweight Algorithm for Breast Cancer Detection

Abstract

As a receptor of breast cancer prognosis and treatment, human epidermal growth factor receptor 2 (HER-2) is closely associated with breast cancer occurrence and progression. Breast cancer lesions are characterized by irregular shapes, small lesion targets, and possible multi-target lesions with overlapping boundaries. Existing algorithms have partly improved the detection accuracy of breast cancer detection model. However, it still suffers from the issues of insufficient detection accuracy and slow detection speed of small and multi-target lesions and requires a huge mass of sample data for iteration during model training, which is demanding on datasets. To address this problem, an improved model is proposed in this paper based on YOLOv10. The model introduces VanillaNet, a lightweight backbone network that can significantly improve detection accuracy by reducing the network depth of the model to equalize detection speed and performance. In addition, the RefConv module is embedded into the C2f structure to further reduce channel redundancy and smooth out lossy situations. In the feature fusion network part, the introduction of a lightweight up-sampling operator content-aware feature reorganization CARAFE module enhances the quality and richness of output features, which effectively improves detection accuracy and speed. The accuracy of the improved model is 98.8%. Thus, the improved model is a significant advantage over mainstream models such as traditional faster region-based convolutional neural network (RCNN), YOLOv5, YOLOv7, YOLOv8, YOLOv10, and YOLOv12.

Improved YOLOv10 for breast cancer detection