Author: Neha Anegondi (United States)
Co-authors: Simon Gao, Verena Steffen, Christina Rabe, Daniela Ferrara, Qi Yang
Purpose
Geographic atrophy (GA) is a major cause of severe visual impairment. Disease heterogeneity, particularly the widely variable GA lesion growth rate, poses a challenge for the clinical development and evaluation of new therapies for GA. Artificial intelligence-based models have the potential to predict disease progression from images. The study aimed to predict baseline GA lesion area and annualised lesion growth rate from baseline fundus autofluorescence (FAF) and/or spectral-domain optical coherence tomography (OCT) images using a multimodal multitask deep learning (DL) approach.
Setting/Venue
Retrospective analysis of baseline imaging data (macular Spectralis [Heidelberg Engineering, Inc., Heidelberg, Germany] OCT volumes [496×1024×49 voxels]; macular 30-degree FAF images [768×768 pixels]) from study eyes of patients with bilateral GA enrolled in natural history and lampalizumab clinical trials (NCT02247479, NCT02247531, NCT02479386).
Methods
For OCT volumes, each B-scan was flattened along Bruch’s membrane (BM), and en face maps averaged over full, sub-BM and above-BM depths were combined as a 3-channel input. GA lesion growth rate (mm2/year) was estimated as slope of a linear fit on all the available measurements of lesion area (mm2, graded by an independent reading centre). The full dataset was split into development (1279 patients/eyes) and holdout (443 patients/eyes) sets. The development set was further split into 5 outer folds. Baseline characteristics were balanced across splits. Three multitask convolutional neural network models were used to simultaneously predict lesion area and growth rate: OCT only, FAF only and multimodal (OCT+FAF). For each, nested cross-validation (CV) was performed on the development set; inner and outer folds were used for hyperparameter tuning/selection and performance estimation, respectively. The models were re-trained on the entire development set using the best hyperparameters from each inner fold, generating 5 models per type. An ensemble of the 5 models was performed to generate holdout set predictions for each type. Performance was evaluated by calculating the in-sample coefficient of determination (R2), defined as the square of Pearson correlation coefficient between true and predicted lesion area and growth rate.
Results
On the development set, the multimodal model had the best CV performance, with mean (SD) R2 of 0.93 (0.03) and 0.52 (0.05) for GA lesion area and GA lesion growth rate predictions, respectively. On the holdout set, the same model showed R2 (bootstrap 95% CI) of 0.94 (0.92, 0.96) for GA lesion area prediction and 0.47 (0.40, 0.54) for GA lesion growth rate prediction. Respective development set mean R2 values for GA lesion area and GA lesion growth rate predictions were 0.91 (0.03) and 0.42 (0.05) for the OCT-only model, and 0.93 (0.03) and 0.48 (0.05) for the model based on FAF images only. On the holdout set, the respective R2 values for GA lesion area and GA lesion growth rate predictions were 0.91 (0.87, 0.95) and 0.36 (0.29, 0.43) for the OCT-only model, and 0.96 (0.95, 0.97) and 0.48 (0.41, 0.55) for the FAF-only model. For comparison, applying a previously developed benchmark model using a simple linear function based on baseline GA lesion area, lesion distance to fovea, lesion contiguity and low-luminance deficit on the same holdout set showed an R2 value of 0.16 (0.10, 0.24) for GA lesion growth rate predictions.
Conlusions
These findings show the feasibility of using baseline FAF and/or OCT images to predict individual GA baseline lesion area and annualised lesion growth rate with a multitask DL approach. The multimodal approach showed slightly improved performance in the development set. However, the performance was comparable with the FAF-only model in the holdout set. Artificial-intelligence-based predictions could be used to inform clinical trial design, implementation, and analysis. Further validation in additional datasets is required to confirm robust performance.
Financial Disclosure
All authors are employees of Genentech, Inc.
Comments
-