Automatic identification of neurodegenerative diseases with 3D point cloud-based analysis using geometric deep learning in OCT retinal images

Neurodegenerative diseases (NDDs) such as Alzheimer’s (AD), Multiple Sclerosis (MS), and Parkinson’s (PD) are becoming increasingly prevalent, requiring reliable biomarkers for early detection and monitoring. Retinal layers, as captured by Optical Coherence Tomography (OCT), offer a promising avenue for automated analysis via deep learning methods. This study explores the use of Geometric Deep Learning (GDL) techniques, which redefine input data, for these prevalent and clinically significant diseases screening using point clouds extracted from Retinal Nerve Fibre Layer (RNFL) and Ganglion Cell Layer (GCL-BM) contours. Three representative GDL architectures were applied to three different analyses: (I) differentiating all NDDs from a control group, (II) separating each NDD from the control group, and (III) performing multi-class classification among the diseases. Optimal point cloud sizes were also investigated. Results showed that in analysis (I), the GDL strategy achieved a high F1-score of 0.92 using only 512 3D points. In analysis (II), with 1,024, 4,096, and 1,024 3D points, it achieved F1-scores of 0.93, 0.94, and 0.97 for AD, MS, and PD, respectively. In analysis (III), multi-class screening reached a F1-score of 0.87. These results demonstrate the effectiveness of using subsampled point clouds for differentiating NDDs and suggest that GDL methods can enhance the efficiency of retinal layer analysis, offering improvements over current state-of-the-art techniques. This highlights the potential of GDL in processing retinal data and advancing NDD detection and classification, with top-performing results obtained using only around 8% of the total 3D points from a sample.