APR-Net: a lightweight multi-scale attention network for oral disease detection

Abstract Multiclass object detection in oral images is important for oral disease screening and computer-aided assessment. Existing lightweight detection models still suffer from insufficient recognition of fine-grained lesions, weak feature representation of low-contrast pathological regions, inadequate multi-scale feature fusion, and missed detections in visually complex oral scenes, which restrict the detection accuracy and recall of categories such as calculus, caries, gingivitis, hypodontia, tooth discoloration, and ulcer. To address these problems, this paper proposes APR-Net , a lightweight multi-scale attention network that integrates and adapts effective downsampling, feature fusion, and attention mechanisms for oral disease detection. Built upon the YOLOv8n framework, APR-Net first introduces an ADown-based lightweight downsampling strategy to reduce fine-grained lesion information loss during feature extraction. Then, RepNCSPELAN4 is adopted to construct a re-parameterized feature pyramid network, enhancing multi-scale feature aggregation for oral diseases of different sizes. Finally, a Partial Self-Attention(PSA) module is embedded after the SPPF layer to strengthen high-level semantic representation and improve the model’s sensitivity to weak-boundary and low-contrast lesion regions. Experimental results show that, compared with the original YOLOv8n, APR-Net improves Precision, Recall, F1-score, mAP50, mAP75, and mAP50–95 by 0.79, 2.85, 2.12, 2.54, 2.07, and 1.62 percentage points, respectively. Meanwhile, APR-Net reduces the number of parameters from 3.01M to 2.79M and increases the overall FPS from 163.14 to 312.16, demonstrating its effectiveness in achieving accurate, lightweight, and real-time multi-class oral disease detection.