人工智能方法诊断黑色素瘤的效果评估:一项回顾性综述
该回顾性研究使用2200张皮肤镜图像,系统比较传统机器学习与深度学习算法在黑色素瘤分类中的表现。结果显示区域生长算法分割最佳(IoU 70.06%,假阳性率17.67%),逻辑回归分类敏感性76.36%、特异性87.04%,Inception V3模型准确率93.74%、敏感性94.36%、特异性85.64%。结论认为传统机器学习在小数据集上表现满意,深度学习在大数据时代潜力巨大。
要点
- 区域生长算法分割性能最佳,IoU为70.06%,假阳性率17.67%。
- 逻辑回归分类敏感性76.36%、特异性87.04%。
- Inception V3深度学习模型准确率93.74%、敏感性94.36%、特异性85.64%。
核心结论
- 传统机器学习在小规模黑色素瘤皮肤镜图像数据集上性能令人满意,深度学习在未来大数据时代潜力巨大。
研究设计与样本
- 回顾性对比研究,使用2200张皮肤镜图像,依次进行图像分割、特征提取和分类,评估传统机器学习算法,并使用近期流行的卷积神经网络框架进行迁移学习训练分类。
主要发现
- 区域生长算法分割性能最佳,交并比为70.06%,假阳性率为17.67%。
- 逻辑回归分类性能较好,敏感性为76.36%,特异性为87.04%。
- Inception V3模型在深度学习算法中表现最佳,准确率为93.74%,敏感性为94.36%,特异性为85.64%。
关键数值与推荐
- 区域生长算法分割交并比70.06%,假阳性率17.67%。
- 逻辑回归分类敏感性76.36%,特异性87.04%。
- Inception V3模型准确率93.74%,敏感性94.36%,特异性85.64%。
临床怎么用
- 在小型皮肤镜数据集上,传统机器学习可提供满意的黑色素瘤分类性能,深度学习模型则展示出更高准确率和敏感性,未来大数据场景下可优先探索深度学习。
局限性
- 实验未对黑色素瘤样本的严重程度进行划分。
- 数据集对于深度学习而言相对较小。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Artificial intelligence methods for the classification of melanoma have been studied extensively. However, few studies compare these methods under the same standards. OBJECTIVE: To seek the best artificial intelligence method for diagnosis of melanoma. METHODS: The contrast test used 2200 dermoscopic images. Image segmentations, feature extractions, and classifications were performed in sequence for evaluation of traditional machine learning algorithms. The recent popular convolutional neural network frameworks were used for transfer learning training classification. RESULTS: The region growing algorithm has the best segmentation performance, with an intersection over union of 70.06% and a false-positive rate of 17.67%. Classification performance was better with logistic regression, with a sensitivity of 76.36% and a specificity of 87.04%. The Inception V3 model (Google, Mountain View, CA) worked best in deep learning algorithms: the accuracy was 93.74%, the sensitivity was 94.36%, and the specificity was 85.64%. LIMITATIONS: There was no division in the severity of melanoma samples used in this experiment. The data set was relatively small for deep learning. CONCLUSION: The performance of traditional machine learning is satisfactory for the small data set of melanoma dermoscopic images, and the potential for deep learning in the future big data era is enormous.
引用
- Cui X, Wei R, Gong L, et al. Assessing the effectiveness of artificial intelligence methods for melanoma: A retrospective review. J Am Acad Dermatol. 2019. PMID: 31255749. doi:10.1016/j.jaad.2019.06.042
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。