Predicting geographic atrophy growth rate from fundus autofluorescence images using deep neural networks
Autor: | Verena Steffen, Qi Yang, Michael Kawczynski, Simon S. Gao, Neha Anegondi |
---|---|
Rok vydání: | 2021 |
Předmět: |
Coefficient of determination
medicine.diagnostic_test business.industry Deep learning Linear model Pattern recognition Pearson product-moment correlation coefficient Confidence interval Lesion symbols.namesake Optical coherence tomography Linear regression symbols medicine Artificial intelligence medicine.symptom business Mathematics |
Zdroj: | Multimodal Biomedical Imaging XVI. |
Popis: | Geographic atrophy (GA) is late-stage dry age-related macular degeneration (AMD). Improved predictors of GA progression would be useful in clinical trial design and may be relevant for clinical practice. The purpose of this study was to accurately predict GA progression over time from baseline fundus autofluorescence (FAF) images (Heidelberg Engineering, Inc., Germany) using deep learning. Study eyes of patients (n = 1312) enrolled in the Lampalizumab trials1, 2 (NCT02479386, NCT02247479, NCT02247531) were included. The dataset was split by patient into training (n = 1047) and holdout sets (n = 265). GA progression, defined as GA lesion growth rate, was derived by a linear fit on all available measurements of GA lesion area (mm2 , measured from manually graded FAF images). The model performance was evaluated using 5-fold cross-validation (CV). Coefficient of determination (R2 ) computed as the square of Pearson correlation coefficient was used as the performance metric. Multiple modeling approaches were implemented, and the best performance was observed using cascade learning. In this approach, pre-trained weights on ImageNet were finetuned to predict GA lesion area followed by further fine-tuning to predict GA growth rate. The 5-folds had an average CV R2 of 0.44, and the holdout showed R2 of 0.50 (95% confidence interval: 0.41 - 0.61). In comparison, a linear model using only baseline GA lesion area in the same holdout showed an R2 of 0.18. Further investigation with visualization techniques might help understand the pathophysiology behind the predictions. The predictions may be improved by combining with imaging modalities like near-infrared and/or optical coherence tomography. |
Databáze: | OpenAIRE |
Externí odkaz: |