[A lightweight multiscale target object detection network for melanoma based on attention mechanism manipulation]

Area of Science:

• Deep learning
• Computer vision
• Medical imaging analysis

Context:

• Melanoma detection remains a critical challenge in dermatological diagnostics.
• Accurate and efficient identification of melanoma targets is essential for timely treatment.
• Existing target detection models may require further optimization for medical image analysis.

Purpose:

• To develop an advanced deep learning target detection model, AM-YOLO.
• To integrate coordinate attention and efficient channel attention mechanisms into the YOLOv5s architecture.
• To enhance the accuracy and efficiency of melanoma recognition.

Summary:

• The AM-YOLO model utilizes Mosaic and MixUp image enhancement techniques.
• It features a modified YOLOv5s backbone and neck network, incorporating coordinate and efficient channel attention mechanisms.
• Comparative experiments demonstrated superior performance of AM-YOLO over other YOLO variants in precision, recall, and mean average precision for melanoma detection.

Impact:

• AM-YOLO significantly improves melanoma recognition accuracy, achieving 92.8% mAP for benign and 87.1% mAP for malignant melanoma.
• The model exhibits a reduced model weight size, indicating improved efficiency.
• This deep learning approach shows promise for clinical application in melanoma target recognition.