A Differential Diagnose Method for Dermoscopy Images

Automated skin lesion detection in dermoscopy images is an essential way to improve the accuracy and efficiency of clinical dermatological diagnosis. Although deep learning-based object detection methods have made dramatic breakthroughs in automatic diagnosis of skin diseases, such tasks are mainly used to identify rare malignant melanoma, but neglect the identification of the most common benign lesions. Meanwhile, the identification accuracy and speed of the existing models are not satisfactory. To address these issues, we propose a residual attention yolov4 (RAYOLOv4) model for skin lesion detection in dermoscopy images, which is composed of a residual backbone, a lightweight Pyramid Attention with Simple Network Backbones(PANet-s), and a yolo head. Specifically, this work replace the original backbone structure with a residual network to improve the model’s ability to learn data features, and a lightweight convolutional block attention model (CBAM) to the existing YOLOv4 detection model to help the model focus more on task-relevant semantic regions. To further improve the robustness of the model, we introduce a series of feasible data enhancement strategies, such as light white balance, keep augmentation methods to expand the data richness. Additionally, an online skin lesion intelligence detection platform is proposed to test the model. We evaluated our RA-YOLOv4 model on the ISIC-skin 2017 dataset. Our results indicate that the proposed RA-YOLOv4 model has outstanding diagnostic performance and interpretability in automated dermoscopic lesion detection scenarios. The final mAP score is 87.1%, which is an encouraging and competitive improvement of about 2.8% compared to the baseline model.