A Deep Learning Based Framework for Diagnosing Multiple Skin Diseases in a Clinical Environment
摘要
- a novel framework based on deep learning
- backbone:eff-b4
- output layer改成14个neuron
- 每个layer group后面接一个auxiliary classifier
- 用t-SNE对image feature可视化
- a dataset that represents the real clinical environment
- 13,603 专家标注的皮肤镜图片
- 14类(扁平苔藓LP,红斑狼疮Rosa,疣VW,痤疮AV,瘢痕KAHS,湿疹和皮炎EAD,皮肤纤维瘤DF,脂溢性皮炎SD,脂溢性角化SK,黑素细胞痣MN,血管瘤Hem,银屑病Pso,暗红色斑PWS,基底细胞癌BCC)
- 精度
- overall acc:0.948
- sensitivity:0.934
- specificity:0.950
- AUC:0.985
- 与280个权威专家比赛:showed a comparable performance level in an 8-class diagnostic task
- a novel framework based on deep learning
方法
- previous work不太适用于实际场景:不符合亚洲人发病率
- database
- this paper调研了北大医学部皮肤病科的database
- from October 2016 to April 2020
- 由同一个技师使用皮肤镜,对着病灶从不同角度,连续拍摄多张
- 2个5年以上经验的专家,结合患者病史,临床表现,皮肤镜特征打标签
- 2人意见不同的时候通过咨询第3人达成一致
- 劣质数据(图像质量、病史不完整、病灶在黏膜/指甲)被排除
- 提取了14 most frequently encountered 常见病,13,603 clinical images from 2,538 patient cases
- 网络
- eff-b4,gradually unfroze
- 7 auxiliary classifiers + 1 final classifiers,element-wise summation
- 无语子
- Comparison with dermatologists
- 280个专家,
- 用独立的测试集:consists of 200 cases with a clinical image and a dermoscopic image,8类(MN, SK, BCC, EAD, SD, Pso, VW and Rosa),每类25
- 模型只用皮肤镜图片
- 都是8选1
一些精度
metrics
dataset
皮肤镜图像示例
总体精度
混淆矩阵