Image Classification of Eye Diseases Using MobileNetV3-Large with CLAHE Contrast Enhancement

Abstract

Background: Deep learning–based image classification of eye diseases has the potential to support initial screening by recognizing visual patterns in images. However, image illumination and contrast can affect the features learned by the model.
Objective: This study aimed to characterize the performance of MobileNetV3-Large on eye image data processed using Contrast Limited Adaptive Histogram Equalization (CLAHE) and analyze class-specific classification errors.
Methods: Images from the Eye Diseases Classification dataset (cataract, diabetic retinopathy, glaucoma, and normal) were resized to 224 × 224 pixels, and CLAHE was applied to the luminance channel of the LAB color space. Data were stratified into training, validation, and test sets (64:16:20). The ImageNet-pretrained model was trained for 30 epochs, and the final-epoch model was evaluated.
Results: Of 844 test images, 768 were correctly classified (accuracy 91.00%; 95% confidence interval: 88.87–92.75%; macro-F1 90.79%). Recall was highest for diabetic retinopathy (99.55%) and lowest for normal images (76.28%), which were most often misclassified as glaucoma. The training curves suggest possible overfitting.
Conclusion: Because no baseline without CLAHE was evaluated, these results characterize the MobileNetV3-Large CLAHE pipeline but do not establish that CLAHE improved classification performance. A non-CLAHE baseline, duplicate-image screening, validation-based checkpoint selection, and external validation are required before any consideration of clinical use.