Moth-flame optimized UNet++ with self-attention for early-stage Alzheimer’s disease prediction using multimodal input

Early-stage identification of Alzheimer’s Disease (AD) is a sizeable challenge to health care globally due to its progressive nature and the fact that there is no available effective treatment. It becomes strategic for practices using interventions meant to halt or reverse cognitive decline if diagnosed early. Recent medical imaging advancements, mainly Positron Emission Tomography (PET) and Magnetic Resonance Imaging (MRI), have greatly unveiled subtle pathological changes associated with this disease. Studies have shown that multimodal neuroimaging can provide crucial information regarding the structural and functional changes in the brain that are associated with AD. However, more research is required to establish sustainable techniques for the detection of AD at all its stages. In this research, a framework of Moth Flame Optimized UNet++ with self-attention is proposed to analyze multi-modal inputs of Structural Magnetic Resonance Imaging (sMRI), PET, and neuropsychological test data for the classification of Alzheimer’s disease ( UNet + + SA –MFO). The framework involves pre-processing the sMRI and PET images to denoise, skull strip, denormalization, next captures complex texture patterns and spatial relationships from both images by placing attention gates at skip connections to ensure reduction of irrelevant features and enhanced localization of significant features. Neuropsychological assessments are passed through fully connected layers of UNet++. Moth Flame Optimization optimizes hyperparameters. Then, a fused feature set is created as a concatenation of all features from multi-modal inputs. Support vector machine (SVM), k-nearest neighbors (k-NN), and Random Forest are used to model, and later weighted stacking ensemble is used to predict the output. The framework is implemented in Python, and evaluation metrics like precision, recall, F1-score, Accuracy, and Area Under the Receiver Operating Characteristic curve (AUC-ROC) are analyzed. UNet++ SA –MFO attains effective 90.5%, 89.9%, 90.2%, 91.8% and 94.1% precision, recall, F1-score, Accuracy and AUC-ROC against existing multi-modal ensemble frameworks. The above findings highlight the potential of enhanced UNet++ augmented with self-attention-based feature extraction and the benefit of integrating innovative optimization for more precise Alzheimer’s Disease diagnosis and classification at the early stage. This proposed framework provides valuable contributions to Alzheimer’s Disease pathology insight, improved diagnostic sensitivity, and ultimately, improved management of this neurological disease.