Skin Lesion Classification via Semi-supervised Learning and Discriminative Loss Function

• Main Authors: YIH-SHIOU LO, 羅翊修
• Other Authors: Chang-Hong Lin
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/awu22w

碩士 === 國立臺灣科技大學 === 電子工程系 === 108 === Deep learning technique has been used to solve computer vision problem recently. One of the most popular applications is medical image classification. Skin lesion is a disease, which can be observed visually and can be cured if professional medical treatments are accepted at an early stage, therefore, developing skin lesion classification system becomes essential. One of the most difficulties for medical images classification problem is to collect images and label them. In this thesis, we used a semi-supervised method to utilize unlabeled images and get a better accuracy. In the experiments, we found out that the cross-entropy loss does not show very discriminative information, so we improved it by adding a penalty loss term. Our loss function shows more information and make the model optimize better than the cross-entropy loss. We combined 5-fold cross-validation and balance weight of different categories with our own loss function and got a mean recall of 0.812 on ISIC2018 dataset.