深度学习在医学影像分析中的应用综述

俞益洲,马杰超,石德君,周振

Application of Deep Learning in Medical Imaging Analysis: A Survey

Yu Yizhou,Ma Jiechao,Shi Dejun,Zhou Zhen

表1 在多种医学影像分析中DL与传统方法和人类的表现对比

Table 1 The performance comparison of DL models with traditional methods and humans on multiple applications in medical imaging

任务和评价指标 数据模态 DL 人工特征方法 人类/专家 DL+人类 数据集 脑肿瘤分割/Dice 核磁 88.0%[91] 79.0%[92] - - 2013 BRATS 转移乳腺癌检测/准确度[93] 病理切片 92.5% - 96.60% 99.50% Camelyon16 视网膜血管分割/准确度 眼底图片 96.0%~97.3%[40] 92.7~94.5% 94.70%[43] - DRIVE/STARE 糖网筛查/AUC 眼底图片 99.0%[94] 87.8%[95] - - Kaggle’s dataset 肺结节筛查/敏感度 CT 95.00%[52] 63.20%[48] - - LUNA16 肝分割/体积重叠误差[89] CT 5.37% 7.73% - - SLIVER07 皮肤癌分类/准确度[13] 皮肤图片 72.10% - 66.00% - Public+Private 乳腺良恶识别/敏感性[74] DBT* 93.00% 85.20% - - Private 肝肿瘤分割/Dice[96] CT 80.1% 75.67%~79.78% - - Private 肠息肉筛查/检出率[90] 结肠镜 96.40% - 7%~53% - Private *DBT:数字乳腺断层合成显像