A multi-resolution hybrid deep feature extraction and ensemble learning framework for OCT-based methamphetamine addiction detection.

Substance addiction is a major health problem that may cause structural and functional changes in the central nervous system. Since the retina is an extension of the brain, it may provide an indirect biological window for observing these changes. Optical Coherence Tomography (OCT) enables non-invasive, micron-level imaging of retinal layers. Although clinical studies have investigated retinal alterations related to substance addiction using OCT, artificial intelligence-based automatic detection studies remain limited. In this study, a hybrid deep learning and machine learning framework is proposed for detecting methamphetamine addiction from OCT images. A new deep feature extraction architecture, ResNet50-OCT, was designed. It was pre-trained on the Stable ImageNet-1K dataset, and multi-layer deep features were extracted from OCT images. These high-dimensional features were reduced using the Multi-Resolution Feature Selection strategy, which combines different feature ranking algorithms and selection ratios. The selected features were classified using Support Vector Machines with different kernels. High-performing model outputs were fused through Threshold Filtered Iterative Majority Voting. The framework was evaluated on 3,859 OCT images from 193 individuals using 10-fold cross-validation and LOSO validation. Accuracy values of 89.43% and 74.07% were obtained, respectively, showing the potential of OCT-based AI for methamphetamine addiction detection.