Deep learning-driven medical image analysis has become a critical auxiliary tool for clinical oral disease diagnosis. However, most existing caries detection methods based on YOLO series models are validated on single-lesion datasets, leading to generally high missed-detection rates and unsatisfactory localization accuracy in complex dental radiographic scenarios where multiple dental conditions and restorations coexist. To address this limitation, this paper proposes YOLO-ASPE, an improved YOLOv11n lightweight framework for tiny-carious-lesion detection. It integrates a P2 high-resolution detection branch, a multi-scale sequential feature fusion module, a lightweight SE channel attention mechanism, and EIoU bounding box regression loss. Experiments were conducted on a public eight-category dental radiographic dataset with caries as the sole detection target. The results show that YOLO-ASPE achieved a Caries AP@0.5 of 68.7%, 10.5 percentage points higher than the baseline model, with a real-time inference speed of 312 FPS. This work provides a lightweight technical reference for the computer-assisted screening of dental caries, although further external validation is still required before clinical application.

Global public safety is increasingly threatened by frequent road traffic accidents, making early accident risk prediction essential for intelligent transportation systems. However, incomplete multimodal data and insufficient spatiotemporal modeling lead to unstable performance in real-world scenarios with sensor failure, data loss, or visual occlusion. To address these issues, we propose the Early Accident Risk Multi-Level Spatiotemporal Robustness Network (EAR-MLSR-Net). Unlike conventional GCN-LSTM or Transformer-based architectures that separately model spatial and temporal dependencies, EAR-MLSR-Net introduces a multi-level spatiotemporal learning paradigm, which jointly models local spatial interactions, regional propagation patterns, and global temporal evolution in a unified hierarchical framework, enabling more structured dependency learning across scales. The proposed EAR-MLSR-Net follows a hierarchical multi-level spatiotemporal learning paradigm that integrates spatial reasoning, cross-modal temporal alignment, and multi-scale temporal dynamics modeling into a unified framework. Experiments conducted on the large-scale CAP-DATA dataset, which contains over 11,727 traffic scenarios and multimodal sensor records, demonstrate that EAR-MLSR-Net achieves an AUC of 0.860, an AP of 0.764, and a TTA_0.5 of 4.252 s. Compared with the strongest baseline, the proposed method achieves a 1.40% improvement in AUC and a 2.10% improvement in AP, while reducing performance degradation by 0.012 under missing-modality conditions.