基于深度学习的色素性皮肤病识别研究  

Research on Pigmented Skin Disease Recognition based on Deep Learning

作　　者：陈澳 CHEN Ao(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydropower Engineering,Three Gorges University,Yichang 443002,Hubei,China;College of Computer and Infomation Technology,Three Gorges University,Yichang 443002,Hubei,China)

机构地区：[1]三峡大学水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《长江信息通信》2024年第3期38-41,共4页Changjiang Information & Communications

摘　　要：由于皮肤镜图片存在着毛发、纹理等方面的干扰,常导致色素性皮肤病识别的误判。为了提高对色素性皮肤病的识别准确率、减少模型的参数量、降低计算量,提出了一种基于MobileViT的色素性皮肤病识别方法。把MobileViT模型作为基础,使用迁移学习训练,并对MobileViT模型作出改进,将MobileViT block的输出融合CBMA注意力机制,对输出使用EfficientNetv2-xl进行知识蒸馏。研究结果表明改进后的算法识别准确率相比原模型提高了7.28%,计算量与参数量也有所降低。并实现了9种色素性皮肤病分类识别界面,为色素性皮肤病在医学辅助诊断方面的研究提供了实验基础。Due to the interference of factors such as hair and texture in dermoscopy images,the identification of pigmented skin diseases often leads to misclassification.To enhance the accuracy of pigmented skin disease recognition,reduce model parameters,and decrease computational complexity,a novel approach based on MobileViT is proposed.The MobileViT model is utilized as the foundation for training,leveraging transfer learning techniques.Further improvements are made by fusing the output of the MobileViT block with the CBMA attention mechanism and applying knowledge distllation using EfficientNetv2-xl.Experimental results demonstrate that the enhanced algorithm achieves a 7.28%incrcase in rccognition accuracy compared to the original model,along with reduced computational complexity and paramcter volume.Moreover,an interface for classifying and identifying nine types of pigmented skin diseases has been developed,providing an experimental basis for research on the medical-assisted diagnosis of pigmented skin diseases.

关 键 词：色素性皮肤病 MobileViT CBMA 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]