Automatic diagnosis of retinoblastoma and quantitative analysis of retinal vascular morphology using deep learning.

Purpose: To develop a convolutional neural network-based artificial intelligence identification diagnostic system for ultra-wide field fundus images of retinoblastoma (RB) and quantify its vascular morphology features. Methods: A retrospective study. 183 ultra-wide field fundus images from 54 patients diagnosed with RB were included, along with 250 images of healthy control. Of these, 277 images were used as the training set, 100 images were used as the validation set, and 56 images were used for the performance test. A deep learning based on DenseNet was used to train the AI auto-diagnostic network for RB. After manual labelling to remove the lesion area, the binary skeleton vessels of the retina were automatically extracted and segmented. Vascular morphological characteristics were subsequently assessed quantitatively using the corresponding mathematical model. Results: The diagnostic performance of the network was evaluated using sensitivity, specificity and F1 score, respectively. In the validation set, the automated diagnostic system achieved 0.98 sensitivity, 0.94 specificity, and 0.96 F1 score. In the test set, the system achieved 0.93 sensitivity, 0.79 specificity, and 0.87 F1 score. In quantitative vascular morphology analysis, compared with the healthy control group, the vascular angle of RB patients was increased, while the vascular density, fractal dimension, and number of vascular branches were decreased (all P < 0.001). Conclusion: In this study, we developed an artificial intelligence system for automatic identification and diagnosis of RB in ultra-wide field fundus imaging, and were able to automatically and quantitatively analyze the retinal vascular morphological features of patients with RB.