Comparative Performance Evaluation of Machine Learning and Hybrid Deep Learning Frameworks for Accurate Detection of Diabetic Retinopathy Haemorrhages

Abstract

The retinopathy that is in diabetic people is among those causes of vision impairment that are mostly experienced globally and the promptness of detecting the haemorrhages of the retina is vital in avoiding serious vision damage. Nevertheless, the manual analysis of the retinal fundus images is time-consuming, subjective, and likely to cause diagnostic variability. Machine learning based detection methods currently have low contrast lesions, structural similarity between haemorrhages and blood vessels, and a weak feature representation that can result in lower detection accuracy with high false positives. Thus, a highly sensitive haemorrhage detection and classification would need a robust and efficient automated system to detect haemorrhage and classify it effectively. The main purpose of this study is a comparative performance analysis of the traditional machine learning models and a proposed hybrid deep learning model on the specific detection of diabetic retinopathy haemorrhages. The suggested solution combines the state-of-the-art pre-processing, candidate extraction, hybrid feature extraction, and deep classification. The combination of handcrafted colour, texture, and shape characteristics and deep features that are extracted with EfficientNet-B7, DenseNet-201 and Swin Transformer networks is used. It trains and assesses these models using the retinal image data to detect the location of haemorrhage and determine the presence of a disease. The experimental results indicate that in both benchmark dataset used in this research and traditional machine learning algorithms can detect objects with accuracy at a range of 8791 and proposed hybrid deep learning framework can detect objects with accuracy of 9598 percent with less false detection and greater resilience to the presence of different image conditions. The better efficiency of the suggested system is proven by comparative analysis conducted to ensure that it has higher detection specificity, sensitivity, and reliability. The conclusion of the study is that the hybrid feature extraction with state-of-the-art deep neural networks is a highly effective and accurate solution to automated haemorrhage detection in diabetic retinopathy to assist in the early diagnosis and the enhancement of the clinical decision-making in an ophthalmic screening system.