[An efficient and lightweight skin pathology detection method based on multi-scale feature fusion using an improved RT-DETR model].

• DOI: 10.12122/j.issn.1673-4254.2025.02.22
• Publication Date: 2025-02-20
• Authors: Yuying Ren; Lingxiao Huang; Fang Du; Xinbo Yao
• Keywords: Feature (linguistics); Sensor Fusion

The presence of multi-scale skin lesion regions and image noise interference limited resources auxiliary diagnostic equipment affect the accuracy disease detection in tasks. To solve these problems, we propose a highly efficient lightweight model using an improved RT-DETR model. A FasterNet was introduced as backbone network FasterNetBlock module parametrically refined. Convolutional Attention Fusion Module (CAFM) used to replace multi-head self-attention mechanism neck enhance ability AIFI-CAFM for capturing global dependencies local detail information. DRB-HSFPN feature pyramid designed Cross-Scale Feature (CCFM) allow integration contextual information across different scales improve semantic expression capacity network. Finally, combining advantages Inner-IoU EIoU, Inner-EIoU original loss function GIOU further model's inference convergence speed. experimental results on HAM10000 dataset showed that model, compared with had increased mAP@50 mAP@50:95 by 4.5% 2.8%, respectively, speed 59.1 frames per second (FPS). parameter count 10.9 M computational load 19.3 GFLOPs, which were reduced 46.0% 67.2% those validating effectiveness proposed SD-DETR significantly improves performance tasks effectively extracting integrating features while reducing both load.