MelaNet:用于皮肤镜图像中黑色素瘤检测的深度密集注意力网络  

MelaNet:A Deep Dense Attention Network for Melanoma Detection in Dermoscopy Images

作　　者：钟昀辛 张朋艺 邓玉林[1] 李晓琼[1] ZHONG Yunxin;ZHANG Pengyi;DENG Yulin;LI Xiaoqiong(School of Life Science,Beijing Institute of Technology,Beijing,100081)

机构地区：[1]北京理工大学生命学院,北京100081

出　　处：《生命科学仪器》2021年第2期40-48,共9页Life Science Instruments

基　　金：北京理工大学重大项目培育基金(1870011162001)。

摘　　要：黑色素瘤是全球范围内一种最致命的皮肤癌。基于皮肤镜图像的黑色素瘤自动化检测对于改善皮肤癌的诊断具有重要意义。因此,本文提出了由专门设计的密集注意力模块构成的深度神经网络MelaNet,通过融合多类别与多标签分类来实现对包括黑色素瘤在内的9种皮肤癌的检测。将年龄、性别及病灶点等元数据作为先验,构建了概率预测模型。在包括8,238张皮肤镜图像的独立测试集上MelaNet取得了先进的检测性能:准确率为86.8%,敏感性为70.8%,以及特异性为86.9%。在ISIC2019国际皮肤病识别挑战赛的元数据排行榜上名列第五,在仅使用ISIC2019数据集的条件下排名第二。MelaNet决策过程的可视化结果展现出了良好的临床相关性,有望用于临床辅助医生进行黑色素瘤等皮肤病变的诊断。Melanoma is one of the deadliest skin cancers in the world.The automated melanoma detection based on dermoscopy images is of great significance for improving the diagnosis of skin cancer.Therefore,this paper proposes a deep neural network,i.e.,MelaNet,which is constructed from specially designed dense attention modules and can detect 9 types of skin lesions including melanoma by fusing multi-class and multi-label classification.By using the metadata such as age,gender,and anatomical sites of skin lesions as priori knowledge,we design a probability prediction model to improve the detection performance.MelaNet has achieved the state-of-the-art detection performance on an independent test set including 8,238 dermoscopy images:Accuracy:86.8%,Sensitivity:70.8%,and Specificity:86.9%.In the metadata leaderboard of International Skin Disease Identification Challenge 2019 (ISIC2019),our MelaNet is the second best method without using additional dataset.The visual explanation for MelaNet’s decision-making process shows high clinical relevance.As a result,our MelaNet has the potential to be used in clinical to assist the clinicians for the diagnosis of skin lesions.

关 键 词：皮肤镜检查 黑色素瘤检测 密集注意力 多类别分类 多标签分类 

分 类 号：R44[医药卫生—诊断学] R751[医药卫生—临床医学]