Enhanced diagnosis of thyroid-associated eye diseases based on deep learning: A novel triplet loss design strategy

Abstract

Thyroid-associated ophthalmopathy (TAO) is an orbital disease that significantly impacts patients’ quality of life. The early diagnosis and treatment of TAO are faced with many difficulties, so some studies have attempted to identify and diagnose TAO at an early stage. However, the diagnostic classification in relevant studies is all based on traditional cross-entropy loss, and the accuracy will decrease under complex conditions with high similarity of eye images. To enhance the precision of TAO diagnosis, this study introduces a data metric method called IP-Triplet, based on triplet loss. Given the data characteristics, we select the DenseNet backbone network for optimization to better extract features from eye images. However, merely modifying the network structure is insufficient. Therefore, inspired by C-Triplet, we use ‘class proxy’ concept to replace the positive and negative samples in the triplet and adjust the distance between the triplets using an enhancement mapper to improve training effectiveness. Finally, this approach is combined with the cross-entropy loss function for mixed training. Our experimental results show that the proposed IP-Triplet loss significantly enhances TAO diagnostic accuracy, achieving a classification accuracy of 95.97 %±0.09, an F1 score of 95.98 %±0.09, and a quadratic weighted kappa score of 96.96 %±0.07. Our model outperforms existing studies on two public datasets, OCT-2017 and OCT-C8, with an accuracy of 99.80 % and 98.18 %, a recall of 99.80 % and 98.18 %, and a precision of 99.80 % and 98.20 %, respectively. Notably, IP-Triplet can be easily integrated into existing CNN models, providing robust support for TAO diagnosis and treatment. The source code is available at https://github.com/lwlwzmu/IP_Triplet_Classification.