ABSTRACT Diabetic retinopathy is a fatal disease worldwide that causes blindness due to high blood sugar levels. An efficient and early identification process is required for preventing patients from blindness. The existing diabetic retinopathy detection models have some significant challenges, such as poor generalization, time consumption, data quality degradation, and inefficient detection. To address these limitations in the context of diabetic retinopathy detection, a novel Deep Neural Transformer RetinoXGBoost framework is proposed that incorporates multiple advanced neural network techniques. For processing the medical images, the APTOS‐2019 and DDR datasets are used as data sources that have multiple retinal images gathered for preprocessing. The preprocessing is crucial for image quality enhancement and ensuring the generalizability of the model. In the extraction phase, the VGG‐19 model is implemented for gathering both low‐level and high‐level visual patterns, and the Swim transformer module is used for the extraction of global and local features. The outputs of these two blocks are concatenated by using the mobile‐former module, and these concatenated features are recalibrated through the attention mechanism. The base‐level classifiers are used in the classification phase, which increases efficiency by optimizing network scaling, enhances computational efficiency by integrating depthwise separable convolution operations, and ensures stable training by aiding residual learning. Multiple base learners are integrated by a stacked ensemble classifier that helps the model in robust decision‐making. For enhancing the interpretability of the proposed model, the explainable artificial intelligence is used in this paper, which highlights critical regions in the retinal images. The experimental validation analysis is performed to demonstrate the superior performance of the proposed model compared to the existing approaches. The proposed model offered an efficient and scalable solution along with a minimum execution time of 120 min and a maximum recognition accuracy of 98.73%.
K et al. (Thu,) studied this question.