Enhancing medical imaging with Ghost-ResNeXt and locust-inspired optimization: A case study on diabetic retinopathy

Abstract

Diabetic Retinopathy (DR) is a global cause of vision impairment, necessitating early detection for effective treatment. Previous works on deep learning architectures like ResNet and DenseNet exhibit promising efficiency in DR detection. However, they're typically computationally expensive and sluggish, rendering them relatively unsuitable for real-time scientific purposes. Also, there are challenges associated with the existing models of displaying normality and the normal distributions to enhance hyperparameters, which often result in the models' generally poor performance. To lower the version complexity and maintain higher accuracy in low-light situations, this examines the cost of the significant GhostNet structure with the former high-throughput ResNeXt model to establish Ghost-ResNeXt. To improve the model performance even more, we propose a new algorithm called Locust-Inspired Optimization, which belongs to swarm intelligence inspired by biological processes. These rules pleasantly tune the model's hyperparameters by emulating the swarm behavior of locusts for the best feature selection and classification. The proposed method entailed training the Ghost-ResNeXt model on a large data set of retinal images. At the same time, the parameters were optimized with the aid of a locust-inspired optimization algorithm. The model design prevents intricate patterns of DR from overshadowing the computational burden while optimally retaining valuable diagnostic information. Experimental outcomes proved that the Ghost-ResNeXT is superior to the conventional architectures in terms of accuracy and pace, which resulted in a remarkable improvement in DR detection accuracy within a comparatively reduced range of parameters. The Locust-Inspired Optimization enhances model stability, achieving higher precision and recall than conventional optimization techniques. This paper provides a novel and useful resource-saving approach to DR identification, which can be applied in ‘live’ diagnostics and telemedicine.