Abstract
Purpose :
Deep learning allows increased accuracy in image recognition and is currently applied for diagnosis in ophthalmic imaging. The accuracy of disease identification and a presentation of the basis for judgment are considered important in diagnostic imaging. An attention map exists that visualizes the attention area used by the model for recognition, which is important to present as the basis for judgment. However, the kind of features the model pays attention to is automatically trained from data, and may differ from that selected by a human expert. Hence, in clinical practice, the basis may differ from the diagnostic basis used by a physician expert.
The purpose of this study is to improve the identification of the attention area and propose a method to incorporate expert knowledge into this area for training the recognition model.
Methods :
This method uses as stepwise approach as follows:
Step1. A disease identification model is trained with the training data set of fundus images, and the test data are recognized. Attention maps for data that apply to the following conditions are acquired:
- Attention maps for incorrect data recognition
- Attention maps for data where the attention area does not correspond to the disease region even though the recognition results are correct
Step2. The attention maps obtained in Step 1 are manually modified to correspond to the disease region.
Step3. The recognition model is fine-tuned to output the correct attention map.
Results :
The accuracy evaluation and subjective evaluation of the attention maps were performed. The recognition accuracy was 96.88% by the proposed method compared to 89.78% by the existing method. Subjective evaluation showed that the attention area other than the disease region decreased, and there was an increase in the agreement in the attention area and the disease region.
Conclusions :
The results of the study indicate that the proposed method improved the agreement between the attention area and the disease region, and it improved recognition accuracy. These outcomes confirm the effectiveness of the proposed method in reflecting the modified attention map for disease evaluation from fundus images.
This is a 2020 ARVO Annual Meeting abstract.