Abstract
Purpose: :
Machine-learning classifiers offer a technique to extract new information from data. This study investigates whether the use of machine-learning classifiers improves the diagnostic accuracy of plus disease diagnosis in retinopathy of prematurity (ROP).
Methods: :
34 digital retinal images were interpreted by 22 ROP experts for presence of plus disease and were allocated for machine classifier training. These images were processed by a computer-based system, Retinal Image MultiScale Analysis (RISA), yielding 6 retinal vessel features: arteriolar integrated curvature (AIC), arteriolar diameter (AD), arteriolar tortuosity index (ATI), venular integrated curvature (VIC), venular diameter (VD), and venular tortuosity index (VTI). Machine-learning classifiers were trained to detect plus disease between eyes with and without plus disease using these six blood vessel features. Six machine-learning classifiers were trained: linear model (LM), generalized linear model (GLM), generalized additive model (GAM), linear discriminant analysis (LDA), support vector machine (SVM), and recursive partitioning and regression tree (RPART). Receiver operating characteristic (ROC) curves were generated and area under the curve (AUC) values were calculated for each classifier.
Results: :
The highest and lowest AUC values among individual features were achieved by VIC (0.853) and AD (0.560), respectively. The highest and lowest AUC values among the classifiers were achieved by the GAM (1.000) and RPART (0.8077), respectively. AUC values for the other classifiers were: LM (0.9634), GLM (0.9670), LDA (0.8993), and SVM (0.9560).
Conclusions: :
Use of machine-learning classifiers of a computer-based image analysis system has potential to diagnose plus disease in ROP with high accuracy.
Keywords: retinopathy of prematurity • clinical (human) or epidemiologic studies: systems/equipment/techniques • imaging/image analysis: clinical