🤖 AI Summary
To address low detection accuracy in X-ray gastric cancer screening—caused by limited annotated samples, low image contrast, and ill-defined lesion boundaries—this paper proposes a lightweight, clinically oriented object detection method. Our approach innovatively embeds boundary-sharpening priors into the loss function, introduces a hard bounding-box supervision mechanism, and designs a fine-grained stochastic data augmentation strategy. Building upon YOLOv4, we integrate Stochastic Label Smoothing, Boundary-Aware Focal Loss, and multi-scale elastic deformation augmentation. Evaluated on a real-world X-ray dataset from a tertiary hospital, our model achieves 92.3% detection accuracy, reduces false positive rate by 37%, and processes each frame in under 80 ms—meeting bedside real-time assistance requirements. To the best of our knowledge, this is the first work to enable end-to-end joint optimization of boundary-aware priors and detection loss, significantly enhancing localization robustness for gastric cancer in small-sample, low-quality radiographic images.
