AUTOMATED QUALITY ASSESSMENT OF OPTIC DISC PHOTOGRAPHS WITH DEEP LEARNING

Abstract

Purpose : To generate and test the performance of a deep convolutional neural network (CNN) with transfer learning for automated assessment of optic disc photograph quality.

Methods : From a large database at a glaucoma tertiary center, the quality of 1741 randomly selected optic disc photos were manually assessed by two clinicians independently. Three levels of quality based on reference images were defined: good, acceptable and poor. This was a heterogeneous database that consisted of both digitized and digital photographs. Preprocessing of each labeled image was performed with a 2400×2400-pixel crop around the optic disc and resizing to 128×128 pixels. Next, we applied seven different image transformations for data augmentation. An optimal three-layer CNN was developed based on a fine-tuned pre-trained VGG-16 network[1]. After image preprocessing, images were split into 80% as a training dataset, while 20% were used as validation dataset. Accuracy, recall, F1-score and a normalized confusion matrix were used to asses model performance. A separate test set of 100 optic disc photographs were used for external validation.


[1] Simonyan, K. and Zisserman, A. (2015) Very Deep Convolutional Networks for Large-Scale Image Recognition. The 3rd International Conference on Learning Representations (ICLR2015).

Results : From the augmented dataset of 13968 images, 7040 (50.5%) were classified as good, 5728 (41.1%) as acceptable, and 1160 (8.4%) as bad quality. The pretrained model showed a weighted accuracy of 91% in the validation set for correct quality classification (Table 1). Per-classes sensitivity was 92% for good quality, 93% for acceptable quality and 81% for poor quality (Figure 1). In the external test set the weighted accuracy was 80% with a weighted recall of 79%.

Conclusions : We propose a CNN with good performance for automatic assessment of the quality of optic disc photographs. This is clinically relevant since evaluation of optic disc photographs for detection of glaucoma is impacted by its quality level, therefore, implementation of an automatic quality assessment could improve the acquisition of these images and ultimately aid in the diagnosis of the disease.

This is a 2020 ARVO Annual Meeting abstract.

 

Classification report per-class, model average and model accuracy.

 

Confusion matrix for performance of the CNN for automatic detection of disc photograph quality. Diagonal squares show the correct proportion of classified images per -class (sensitivity).