Abstract
Purpose :
Multiple strategies exist to identify disease from fundus photography as well as localize and differentiate red from yellow/white lesions. However, the ability to computationally differentiate between subtypes of bright lesions remains limited. This study combined three fundus photograph datasets for transfer learning with the VGG16 algorithm to develop a three-class classifier for drusen, hard exudates, and hemorrhages.
Methods :
Regions of single and clustered lesions were identified from the SUSTech-SYSU exudate dataset, Indian Diabetic Retinopathy Image Dataset (IDRiD), and Fundus Image Vessel Segmentation (FIVES) dataset (Figure 1). All selections were verified by an independent observer.
All layers of the convolutional base of the ImageNet trained VGG16 algorithm were fixed except for layers 5-1, 5-2, and 5-3. A flatten layer followed by fully connected layers with 50, 20, and 3 nodes respectively were appended. Data was rescaled to 150x150 using nearest neighbor interpolation and augmented by vertical and horizontal flip, width and height shift within 10%, and rotation within 30°. Adam optimization was used at a learning rate of 0.01 for 13 epochs and 0.0001 for 10 epochs. Additionally, a modified algorithm with ImageNet weights substituted in the convolutional but not fully connected layers was evaluated.
Results :
Detailed validation and testing results are reported in Table 1. Resetting convolutional layer weights to ImageNet trained values decreased algorithm accuracy from 0.9370 to 0.3386 on validation data, equivalent to that of a random classifier.
Conclusions :
Trained and tested on a slightly larger dataset than previous studies, this model had substantially better performance on classification of bright retinal lesions. Furthermore, its classification of hemorrhages trained on one dataset (IDRiD) was generalizable to a different dataset (SUSTech-SYSU). This approach of transfer learning with convolutional neural networks easily scales for identification of more lesion types.
Additionally, the trained convolutional layers were imperative to model function. Thus, we are now developing bounding box object detectors that build on convolutional models pre-trained on classification of isolated lesions with the goal of improving training time and performance in images with overlapping findings.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.