Abstract
Purpose :
Automated machine learning (AutoML) is a new area in artificial intelligence that automates important components in the machine learning design pipeline. It has not been broadly used for retinal disease detection from fundus images. This study aimed to assess the discriminative performance of AutoML in classifying retinal pathologies from ultra-widefield (UWF) fundus pseudocolor images.
Methods :
Using Google Cloud AutoML Vision, we developed a deep learning classification model to detect normal fundi, retinal vein occlusion (RVO), retinitis pigmentosa (RP), and retinal detachment (RD) from UWF pseudocolor images. We selected those diseases for the presence of retinal hemorrhage in RVO, retinal pigment epithelial changes in RP, and retinal elevation in RD. We used an open-access dataset provided by the Tsukazaki Optos Public Project (13,047 images) to curate a smaller dataset of training and validation images. Two ophthalmologists (raters) independently graded the images according to consensus criteria. Selected images were then uploaded to the online platform for testing and training through the graphical interface and without any coding. Cohen's kappa coefficient (κ) was used to assess the inter-rater agreement during data labeling. The performance of the AutoML model is reported using the area under the precision-recall curve (AuPRC), precision, and recall.
Results :
A total of 2,311 images were relabeled. There was almost perfect agreement between the two raters for normal (κ = 0.82), RVO (κ = 0.85) and RD (κ = 0.82) images, and substantial agreement for RP images (κ = 0.75). The final dataset included 1,355 images (472 normal, 311 RVO, 187 RP, and 385 RD). The overall AuPRC was 0.95, and the precision and recall were both 91.85%. The per-label precision and recall were as follows: normal (88.46% and 97.87%), RVO (96.43% and 87.1%), RP (100% and 78.95%) and RD (90% and 94.74%).
Conclusions :
We demonstrate the feasibility of using AutoML by ophthalmologists without coding experience to create a deep learning disease classification model from UWF images. The model can detect RVO, RP, and RD (three retinal diseases with distinctive clinical features) with excellent precision and recall. With that being established, further research aimed at developing more challenging classification tasks is being carried out by our group (e.g. disease types and stages).
This is a 2021 ARVO Annual Meeting abstract.