A Method For Dermatoscopic Image Differential Diagnosis Based On Object Detection Mechanism

The ’moles’ on the skin can be classified into many categories,such as benign melanocytic nevi,malignant melanoma,and actinic keratosis,among others.Due to the similarities in appearance between malignant melanoma,actinic keratosis,and common nevi,manual skin examination with a dermatoscope can lead to missed or misdiagnosed cases.Convolutional neural networks(CNNs)can automatically analyze skin lesion images and provide more objective and reliable computer-aided diagnosis results by learning from a large amount of data.Therefore,exploring the use of such models to assist doctors in improving diagnostic accuracy is of great significance.Currently,there are several challenges in using CNNs for automated analysis of skin lesion images.Firstly,common classification algorithms used for localized lesion diagnosis lack precise lesion localization,resulting in ambiguous outcomes for malignant melanoma.Secondly,in the three-classification task of skin lesion images,the high intra-class heterogeneity and low inter-class similarity make it difficult to localize lesion information.This paper addresses the aforementioned problems by using the Multi-scale Fusion Single Shot Multibox Detector(MFSSD)and the Real-time Attention You Only Look Once version 4(RA-YOLOv4)to achieve binary and ternary classification of dermatoscope images.Furthermore,the effectiveness of the models in practical diagnostic scenarios is verified by building a corresponding computer-aided diagnosis system.Specifically,this research includes the following 3 aspects:(1)Lightweight MFSSD for Binary Classification of Dermatoscope ImagesTo address the problem of imprecise lesion localization in the binary classification task,a lightweight MFSSD object detection solution is proposed.This method is based on the SSD model and introduces a feature pyramid structure to enhance the model’s fusion capability for features of different scales.Experimental results show that this model improves detection performance by 4.87% compared to the original detection model on the publicly available ISIC dataset.Compared to existing classification approaches,this model reduces the model size by102 M while maintaining similar classification performance.This indicates that the proposed method outperforms the original detection algorithm in terms of both detection performance and model size.(2)High-precision RA-YOLOv4 Detector for Classification of Dermatoscope ImagesTo address the difficulty of lesion localization in the ternary classification task,a highprecision RA-YOLOv4 object detection solution based on YOLOv4 is proposed.This solution replaces the model backbone of YOLOv4 with a residual network and incorporates attention mechanisms to diagnose lesions more accurately.Experimental results show that compared to the original model,RA-YOLOv4 achieves a 4.18% improvement in mean average precision(map)on the publicly available ISIC dataset.Compared to existing classification approaches,this model achieves the highest F1 score(86.67%),surpassing the best existing classifier by0.82%.(3)Design and Implementation of A Skin Lesion Detection SystemTo address the current lack of automatic computer-aided diagnosis systems for skin lesion detection,an automatic detection and computer-aided diagnosis system for mole images is proposed.This system consists of a Python-based backend lesion detection system and an HTML-based frontend online lesion detection system.Practical examples of the system demonstrate its application value and the feasibility of the proposed algorithms in real-world diagnostic scenarios.