<p>Due to the similarity of medical image, the accuracy of medical image diagnosis is slightly lower. To improve the accuracy of medical diagnosis and classification, a metric learning method based on triple loss has been proposed for medical image classification. Firstly, the distance between negative groups has been introduced into the triplet loss to accelerate the decay rate and improve the training efficiency. Secondly, angle loss has been integrated into the loss function to correct the gradient direction and accelerate the convergence of the training process. Thirdly, an adaptive margin calculation strategy has been incorporated into the triplet loss to represent the distance measurement more accurately. Finally, the feature vectors have been fed to SVM for classification. The datasets of dermatosis and cervical cancer have been introduced into the experiment, the experimental results show that the effectiveness and the accuracy of the proposed method have also reached the level of existing technologies, which can be applied to various clinical diagnoses in the future.</p>