advertisement
Editors Selection IGR 24-3
Anatomical Structures: Computer-Assisted Fundus Assessment
Comment by Kouros Nouri-Mahdavi on:
75175 Hybrid Deep Learning on Single Wide-field Optical Coherence tomography Scans Accurately Classifies Glaucoma Suspects, Muhammad H; Fuchs TJ; De Cuir N et al., Journal of Glaucoma, 2017; 26: 1086-1094
Use of deep learning methods has become a hot topic in the field of Ophthalmology. Muhammad et al. tested a hybrid deep learning (HDL) approach for detection of eyes with definite glaucoma from healthy glaucoma suspects using a combination of a pre-trained Convolutional Neural Networks (CNNs) for feature extraction followed by a random forest model for classification.
One hundred and two eyes from 130 eyes from a previous study were used to define glaucomatous vs. normal eyes. During this process, eyes that were harder to classify, i.e., those that were forced-choice into glaucoma or normal groups by the expert reviewers in the original study, were excluded. The gold standard was presence or lack of glaucoma as determined by two glaucoma specialists based on 24-2 and 10-2 VFs, disc photos, patient chart information, and the single-page OCT report as previously published by Hood et al. containing retinal nerve fiber layer (RNFL)n and macular retinal ganglion cell/inner plexiform layer (RGC+) data. The input to HDL was a single wide-field OCT image (12x9 mm, 256 horizontal B-scans each consisting of 512 A-scans). HDL's performance was compared to that of established OCT parameters (average RNFL, quadrant and clock hour sector thickness) or 10-2 or 24-2 metrics. Outcomes of interest were area under the ROC curves (AUC) and model accuracy, which was seemingly the average of sensitivity and specificity values.
Overall, the HDL's performance using RNFL thickness or RNFL probability maps was excellent with an AUC of 0.970 for the RNFL thickness. The best OCT metric misclassified 13 (10%) eyes and the best 24-2 (abnormal GHT or PSD) and 10-2 VF metrics (abnormal MD, PSD, or cluster of abnormal points) misclassified 20 eyes (15%). The inadequate performance of VF metrics is not unexpected given the fact that it is not uncommon to see evidence of RNFL or RGC+ damage in the absence of clear cut evidence for VF loss.
The results are very promising since HDL could be easily incorporated in clinic and could basically elevate the performance of the average clinician to that of glaucoma experts. One has to keep in mind that the HDL's performance could have been overestimated as the controversial cases were removed from the original database. The investigators beautifully demonstrated how exploration of outliers (false positive and negative cases for the HDL classifier) could help better understand the behavior of HDLs. A limitation of this approach was the use of a pre-trained CNN (AlexNet). As mentioned by the investigators, further training of the CNN can be incorporated in the future iterations of HDL algorithms. Therefore, enhancing HDL's performance can be accomplished by not only better training of the neural networks but also by defining a priori anatomical variations and caveats that can lead to false negative or positive results.
Comments
The comment section on the IGR website is restricted to WGA#One members only. Please log-in through your WGA#One account to continue.Log-in through WGA#One
