小样本域自适应的皮肤病临床影像识别方法

Few-Shot Domain Adaptation for Identification of Clinical Image in Dermatology

针对公开数据集训练所得模型无法直接应用于临床上不同设备的辅助诊断,而临床获取的数据又缺少足够人力进行标注的问题,提出了一种面向皮肤病临床影像识别的小样本域自适应方法。以ISIC皮肤病公开数据集作为标签已知的源域,以实际临床采集的数据作为待识别的目标域,通过医生对极少量临床数据进行标注,建立由卷积神经网络实现的特征提取器和分类器,构建小样本域自适应模型。引入最大相关熵准则来提高识别模型的精度和泛化能力,在每类只有少量带标签目标域样本的情况下,通过交替最大最小化条件熵,在提取区别性特征的同时减小不同域之间的分布差距,提高了分类器在新域上的准确率,实现了模型的跨域迁移。对所提方法在日光性角化病和脂溢性角化病分类问题上进行了实验验证,结果表明:相比于非域自适应方法,所提方法克服了不同采集设备造成的数据分布差异问题,取得了更高的识别准确率;相比于无监督域自适应方法,所提方法通过加入极少量标注的临床数据实现了域自适应,识别准确率为93.94%。

Aiming at the problem that the model trained by public data set cannot be directly applied to the auxiliary diagnosis of different clinical devices and there is not sufficient manpower to label the clinical data,few-shot domain adaptation for the identification of clinical image in dermatology is proposed.The ISIC dermatological public data set is taken as the source domain with known label,and the actual clinical dataset is taken as the target domain to be predicted,small amount of clinical data are marked by the doctor to train few-shot domain adaptive model of feature extractor and classifier implemented by convolutional neural network.The maximum correntropy criterion is adopted to improve the accuracy and generalization ability of the recognition model.If there are only a small number of labeled target samples in each class,the distribution gap between different domains is reduced while extracting discriminative features by alternating maximum and minimum conditional entropy.The accuracy of the classifier in the new domain is improved,and the model is transferred across domains.The proposed method is experimentally verified in the classification of actinic keratosis and seborrheic keratosis.Compared with the non-domain adaptive method,the proposed method solves the difficulty of difference in data distribution caused by different collection devices and achieves higher recognition accuracy.Compared with the unsupervised domain adaptive method,the proposed method achieves domain adaptation by adding a small amount of labeled clinical data,and the recognition accuracy rate is 93.94%.