Purchase this article with an account.
Helen Jane Kuht, Garima Nishad, Sharon S-H Wang, Gail Maconachie, Viral Sheth, Zhanhan Tu, Michael Hisaund, Rebecca J McLean, Ravi Purohit, Seema Teli, Frank A Proudlock, Yu-Dong Zhang, Irene Gottlob, Girish Varma, Mervyn G Thomas; A machine learning solution to predict foveal development and visual prognosis in retinal developmental disorders. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2739.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Detecting and grading abnormal retinal development has important diagnostic and prognostic implications in children. The foveal hypoplasia (FH) grading system is based on the stage at which retinal development ceases, and can be described as grade 1 to 4 FH and atypical FH. Correctly identifying the degree of arrested retinal development requires an understanding of the retinal developmental sequence and thus can be challenging for the non-expert. We therefore proposed the development of the first artificial intelligence (AI) based system to accurately identify and grade FH using optical coherence tomography (OCT).
A representative sample (n=5078) of paediatric OCT scans demonstrating varying degrees of arrested retinal development (normal, grade 1-4 FH and atypical FH) were obtained from the Leicester paediatric OCT database. Foveal scans were acquired from table mounted OCT (n=3037) (Copernicus, Optopol Technology S.A., Poland) and handheld OCT (n=2041) (Envisu 2300; Leica Microsystems, Germany) devices to ensure a high-performing, device agnostic system. The foveal B-scans were extracted, annotated, and segmented. A high-yield training dataset (n=3555) was inputted through a customised convolutional neural network (CNN) (Resnet50) for the training stage. Following the training and fine-tuning of the customised Resnet50, a validation stage to test the accuracy of the model was implemented. The foveal scans in this stage were new and unseen scans to the CNN.
Our binary classification (normal and abnormal foveal morphology) and six-point classification (normal, grade 1-4 FH and atypical FH) achieved 98.1% and 95.0% validation accuracy, respectively.
We have demonstrated, for the first time, the proof-of-concept for the use of a device agnostic, automated AI system in paediatric OCT interpretation. Our system can help to eliminate inter-examiner variability and augment the clinical pathway by increasing time efficiency during busy clinics. The introduction of OCT to routine clinical assessment is imminent. Therefore, our AI system provides a strong foundation for the development of a real-time, frontline diagnostic tool for retinal developmental disorders.
This is a 2021 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only