A Lightweight Deep Learning Model for Retinopathy of Prematurity Classification in eHealth Applications

Retinopathy of Prematurity (ROP) is a vision-threatening condition in premature infants requiring timely and accurate diagnosis to prevent blindness. While electronic health (eHealth) technologies promise to improve neonatal care, automating ROP diagnosis faces challenges such as limited labeled datasets, architectural complexity, and high computational demands. This study introduces LightEyeNet, a lightweight deep-learning architecture optimized for eHealth applications in ROP severity classification. By integrating DenseNet121 and a channel-wise residual attention network block, LightEyeNet enhances diagnostic accuracy and efficiency. Explainable AI techniques, including Grad-CAM and LIME, further improve transparency and clinical interpretability. LightEyeNet achieves 96.28% testing accuracy, outperforming state-of-the-art pre-trained networks, including DenseNet201 (95.78%, + 0.5%), Inception-V3 (93.80%, + 2.48%), Xception (94.54%, + 1.74%), and EfficientDense (87.10%, + 9.18%). Furthermore, LightEyeNet is the most compact architecture among these, with a size of 2.45 MB, compared to Efficient-Dense (6.88 MB), Inception-V3 (3.56 MB), Xception (21.48 MB), and DenseNet201 (5.05 MB). With a specificity of 0.99, a sensitivity of 0.95, and an AUC score of 0.99 across five ROP severity classes, LightEyeNet demonstrates a balance of superior performance and efficiency.