SILVA, W. M.; SILVA, Wendson Magalhães da.
Abstract:
Ophthalmic diseases, such as cataracts, glaucoma, and diabetic retinopathy, pose a significant challenge to public health, with the potential to cause vision loss. However, the majority of these cases could be prevented or treated if diagnosed early. In this context, fundus imaging emerges as an effective, fast and non-invasive diagnosis tool. Manual interpretation of ophthalmic images is repetitive and prone to error. Thus, computational systems can be used to assist professionals in automated screening, reducing time, errors, and effort in disease analysis. Deep learning systems have proven effective in this context, however, their lack of transparency has been a challenge for clinical adoption, highlighting the importance of explainability in machine learning models. This study contributes to advancing the understanding and interpretation of deep learning models in the field of ocular health, aiming to improve the diagnosis and treatment of ophthalmic conditions. It compares the LIME and Grad-CAM techniques applied to different architectures of convolutional neural networks (CNNs) trained to classify ophthalmic conditions from fundus images. The results indicate that the VGG16 model stands out, achieving an accuracy of 93.17% in training and 87.16% in validation. Additionally, explainability techniques, though distinct in approach, identified nearly the same regions of interest in ophthalmic images. Nevertheless, despite limitations such as the randomness of LIME and the need for adjustments in Grad-CAM, LIME highlighted critical areas more subtly, while Grad-CAM provided more direct and intuitive visual representations.
