advertisement
Editors Selection IGR 20-4
Clinical Examination Methods: Artificial Intelligence applications I
Comment by Gustavo de Moraes & Bruna Melchior Silva on:
82092 Accurate prediction of glaucoma from colour fundus images with a convolutional neural network that relies on active and transfer learning, Hemelings R; Elen B; Barbosa-Breda J et al., Acta Ophthalmologica, 2020; 98: e94-e100
The lack of symptoms in early glaucoma can lead to late diagnosis; therefore the development of a cost-effective screening method may have an important role in preventing vision loss due to glaucoma.
The objectives of this study were (1) to develop and validate a deep learning glaucoma classifier based on color fundus images of glaucoma patients compared to the clinical diagnosis based on full ophthalmologic examination, tonometry, visual fields, and OCT or HRT; (2) to explore the added value of active learning on top of deep learning for automated glaucoma detection; and (3) to inspect the trained model's decision process using interpretable heatmaps.
A total of 7,038 fundus images passed through quality assessment and were allocated to training (70%), validation (10%), and testing sets (20%). Data augmentation was randomly generated to increase the number of images used to train the convolutional neural network (CNN). Area under the receiver operating characteristic curve (AUC) was selected as main performance metric, with specificity and sensitivity also reported. The CNN was backed up by the analysis of false positives and false negatives by two ophthalmologists in a blind fashion.
Two trained glaucoma specialists analyzed more than 500 heatmaps and indicated a recurrent pattern in the inferotemporal and superotemporal zones neighboring the optic nerve head (ONH), providing novel insights into the decision-making process of the trained deep learning glaucoma classifier through these maps, suggesting that regions outside the ONH could be valuable in this analysis.
This study yielded a deep learning-based glaucoma classifier that achieved an AUC of 0.995 for patient referral with a 60% decrease in labelling cost through the combina-tion of transfer learning, careful data augmentation, and uncertainty sampling. One major strength was the topographical analysis of individual heatmaps to better understand the reasons for errors.
In combination, these findings could help advance the field of artificial intelligence applied to ophthalmology, especially when insufficient data are available.
