小样本域自适应的皮肤病临床影像识别方法

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%.