Abstract
Purpose :
To establish a glaucoma diagnosis system based on deep learning algorithm，which suitable for asian race，and verify it's performance in different medical environments. Verify the interpretability of the conclusions of the algorithm by analysis the visualization result.
Methods :
The training datasets was included 123,398 images,15,100 fundus images of possible glaucoma and 108,298 images of normal eyes. These images were derived from Chinese Glaucoma Study Alliance (CGSA) covering almost the whole of China.All the images were graded by three layers of trained graders of increasing expertise for glaucomatous retinopathy.The validation datasets consists of four sources,including 51448 images.In addition to the local validation dataset we have also set three additional datasets(clinic-based data set, multi-quality data set and population-based data set.)
We visualized the contributions of different regions in fundus images to predicting glaucoma of the algoritm.We resized each original fundus image into a 360x360 RGB image. Then, a 60x60 gray block was used to slide through the fundus image (with a stride of 10 pixels), alongside both horizontal and vertical axes. Consequently, the fundus image generates 961 (=31x31) visualization testing images, each of which has a 60x60 gray block at different position, respectively. The prediction probability output by the algoritm refers to the value of the visualization heat map at the corresponding position.
Results :
The local validation dataset consisted of 21654 images (mean age, 54.4 years; 62.2%women); and the AUC of GD-CNN for referable glaucoma was 0.994, sensitivity was 0.965(95%CI, 0.956-0.973), and specificity was 0.956 (95%CI, 0.953-0.959). In the three additional datasets, AUC range was 0.823 to 0.998 (n = 29794images).
1000 images performed on the visualization layer of the algorithm. The regions of interest in 91.8% of images were identified to have the greatest contribution to the neural network’s assignment and these features are what ophthalmologists use to make a diagnosis.
Conclusions :
The glaucoma diagnosis system performs well in the validation of multiple source validation sets. The visualization results confirm that the algorithm has certain clinical interpretability for the judgment of the fundus image.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.