Detection of Plus Disease in Retinopathy of Prematurity Using Deep Learning Models: Evaluating Vision Transformers and ResNet Architectures

To evaluate the performance of Vision Transformer (ViT) and ResNet-50 in detecting Plus Disease (PD) on fundus color images and vascular segmented mask images of Retinopathy of Prematurity (ROP) patients. A dataset consisting of 1205 fundus color images of ROP patients was extracted from the registry of a leading Research Hospital in Istanbul. Using these fundus images, a second dataset of vascular segmented mask images was created with a U-net segmentation model. The performance of ViT and ResNet models in detecting Plus Disease was evaluated on both sets of images. External validation of the model performances was carried out using a public domain dataset. For fundus color images, ViT models performed better than ResNet in terms of accuracy (96.9% vs. 91.5%), precision (97.1% vs. 85.5%), and F1 score (96.9% vs. 92.2%). However, ResNet had a better recall rate (100% vs. 96.9%). For segmented images, all performance measures were better with ResNet than ViT: accuracy (91.5% vs. 82.7%), precision (85.5% vs. 82.9%), recall (100% vs. 92.3%), F1 scores (92.2% vs. 82.6%), and AUC (99.8% vs. 88.6%). The strong performance of the ViT on fundus color images highlights its potential as a promising model for PD detection. However, its higher computational cost suggests that further optimization will be needed in future research. ResNet-50, with its solid overall performance and perfect recall rate—ensuring no false negatives—appears to be an optimal choice for PD detection. Additionally, vascular segmentation did not provide any enhancement to the model performances.