Abstract
Purpose :
Deep learning (DL) has achieved groundbreaking results in automated image classification over the last decade. The application of DL to medical images promises increased diagnostic accuracy and consistency, reduced clinical burden, and improved workflow. A DL model was trained and tested on preoperative optical coherence tomography (OCT) images on two sets of patients: those diagnosed with full thickness macular holes (FTMH), and those with Epiretinal membrane (ERM) only.
Methods :
This study includes OCT images of 73 eyes from 69 patients with FTMH (51 Females, 18 Males, ages 52-84, mean: 69.6, SD: 6.4) with confirmation of diagnosis and image quality assessed by two trained readers. The remainder of the dataset consisted of scans from patients diagnosed with Epiretinal membrane (ERM), comprising 297 eyes from 287 patients (144 female, 143 male, ages 23-93, mean: 70.5, SD: 8.6). Severe comorbid ocular conditions and/or ocular complications of systemic diseases were excluded. Best corrected visual acuity was the primary efficacy variable analyzed pre- (<2 months prior) and post-operative (6 months after). Factors such as age, gender, surgeon, lens status (phakic or pseudophakic), intraocular pressure, medical, and surgical history were also analyzed. Open-source software (TensorFlow) was used to build the model, based on a convolutional neural network (CNN) architecture. A larger set of retinal OCT images was used to pre-train CNN weights using a self-supervised approach. The final model was fine-tuned on the training set itself. A test set comprising 12 FTMH and 42 ERM-only eyes was set aside for model evaluation post-training.
Results :
The receiver operating characteristic for FTMH prediction showed 0.994 area under curve (p < 0.001), with 11 of 12 FTMH+ scans and 40 of 42 ERM scans accurately diagnosed at the standard threshold of 0.5.
Conclusions :
Our results demonstrate the potential of DL algorithms for computer-assisted screening of FTMH, improving clinical workflow and efficiency. Saved time can be used towards a higher level of patient care or increasing the number of patients seen. Self-supervised pre-training followed by fine-tuning on the target task, may be applied to other clinically relevant image-based diagnoses in ophthalmology and other medical disciplines.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.