Abstract
Purpose :
To provide real-time, automatic Diabetic Retinopathy (DR) screening using retinal images from patients photographed at diabetes clinics. This new technique integrates amplitude modulation-frequency modulation (AM-FM) features and deep learning (DL) methodology.
Methods :
VisionQuest has implemented at multiple diabetes clinics in Mexico an auto-screening system based on AM-FM, which has been improved further through DL techniques, as follows:
(1) Preprocessing and Data Augmentation: A training set of 136,100 images and an independent validation set were standardized for size and lighting. (2) Transfer Learning: Using Google’s Inception-v3 network, DL experiments were performed with different training sets, number of iterations, learning rates, and batch sizes. (3) AM-FM Feature-based model: The features from AM-FM were correlated to anatomical and pathological features associated with DR.
(4) Algorithm validation was performed on retinal images from 680 diabetic patients collected over 3 years at a network of diabetic clinics in Monterrey, Mexico. The images included a wide range of image quality and pathology. None were used for training, and none were excluded from analyses for any reason.
Results :
Sensitivity and specificity for detection of severe non-proliferative retinopathy, as compared to the reading of certified retinal graders are reported. For the AM-FM-only algorithm, the sensitivity/specificity was 91%/84%. For DL-only, sensitivity/specificity was 91%/86%. Combining the outputs of the algorithms yielded an equal error sensitivity/specificity of 91%/91%, with a 95% area under the curve.
Conclusions :
The combination of DL and AM-FM features increased specificity, without sacrificing sensitivity. From our tests on real clinical cases with varied image quality and pathology it is evident that DL image classification and the top-down spatial method of AM-FM are both useful in determining levels of pathology in retinal images. The combination of the two approaches achieves significantly improved results.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.