Content-Based Retinal Image Analysis: Exploring the Impact of Similarity Measures and Machine Learning on LTP Features

Diabetic Retinopathy (DR) is a retinal condition resulting in damage to blood vessels within the eye, serving as a leading cause of vision impairment or blindness when not addressed. Manual identification of diabetic retinopathy is labor-intensive and susceptible to human error due to the intricate nature of the eye's structure. This work offers a comprehensive method that uses image pre-processing techniques, local ternary pattern (LTP) features, machine learning and different similarity measures to improve content based retinal image analysis and retrieval from fundus images. With an emphasis on feature extraction, subtle texture details were extracted from retinal fundus images using Local Ternary Patterns (LTP) with different radius values, specifically concentrating on radius 1, 2, and 3. This study includes two parts: first, it assesses machine learning algorithms for the classification of diabetic retinopathy. Of these, Random Forest performed the best, with an accuracy of 92.77%. Second, the research works on retinal image retrieval and compares various LTP variants' abilities to different retrieval tasks. To evaluate model performance for particular disease classes, class-wise image retrieval was carried out, providing information on individual class precision. Several similarity measures were investigated for retinal image retrieval, demonstrating the effectiveness of Cosine Similarity in achieving better retrieval results.