Abstract
Purpose :
Automatic diabetic retinopathy (DR) diagnosis is of great value for the timely treatment, and promising performances are achieved by deep learning models recently. However, due to domain shift issue, the performance of a well trained model might degrade when deployed to a new condition, e.g., a new hospital or camera.The aim of this study is to improve the performance of the model in the case of domain shift. The proposed method is demonstrated effective to improve the performance for a novel domain at a low cost, i.e., only a few unannotated samples from the target domain is required.
Methods :
The formulation of our method follows (Wang et al.,2021). In particular, we have the parameters of a deep learning model (ResNeXt) trained by large-scaled datasets from source domain. A few (about 1%) unlabeled samples from the target domain are available for model adaptation. To address the domain shift issue, we introduce two techniques: 1) fine-tuning the original model by entropy minimization; 2) test-time data augmentation. Specifically, we measure the confidence of a model by the entropy of its predictions and use it as an adaptation objective. Then, during the inference phase, we apply data augmentation on an input sample and generate K prediction scores. Finally, the multiple predictions are merged by the temperature-shapern method.
Results :
The source domain we used is pulic EyePacs dataset (50k samples belonging to 5 DR grades), while the target domain set is from a new clinic, including 3k samples for adaptation and 465 annotated samples for evaluation. Our experimental results show that our method achieves the best results in all cases, achieving 0.6446 Kappa, 0.9262 AUC and 0.6887 ACC. As shown in the attached table, the proposed method brings substantial improvement for baseline model, i.e., the improvment in ACC is over 5%. In comparison with the previous method, Tent, our method presents consistent improvement in all metrics, while Tent decreases the Kappa score, which is considered as the most important metric in DR grading.
Conclusions :
This work studies the domain adaptation for deep learning based automatic DR grading. A method based on entropy minimization adaptation and test-time data augmentation is proposed. With only 1% unlabeled samples from the target domain, our method achieves a substantial performance boost, showing promising application value.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.