Explainable AI-Based Medical Image Segmentation and Radiomics for Brain Disorder Classification over Structural MRI

Neuropsychiatric disorders including attention-deficit/hyperactivity disorder (ADHD), bipolar disorder (BD), and schizophrenia (SZ) exhibit significant clinical heterogeneity and overlapping symptom profiles, which severely limits the diagnostic specificity of conventional symptom-based assessments. While structural magnetic resonance imaging (sMRI) provides objective insights into anatomical brain alterations, existing computational frameworks often rely on single-plane or binary patient-control analyses, lack anatomical interpretability, or risk data leakage through inconsistent validation protocols. To address these challenges, this study proposes a segmentation-guided AI framework for medical image processing and recognition that fuses FastSurfer-derived regional handcrafted features encompassing Fourier shape descriptors, two-dimensional wavelet and second-order local directional gradient pattern (LDGP) textures, and ROI-filtered XFeat keypoints with complementary multi-view (axial, sagittal, coronal) EfficientNetB0 deep representations extracted from whole-brain T1-weighted sMRI. The segmentation stage provides anatomically defined regions for radiomic feature extraction, while explainable representations maintain anatomical traceability. A key contribution of this work is its dual evaluation strategy within a unified structural feature space: it addresses both broad transdiagnostic four-class classification (Healthy Control (HC) vs. ADHD vs. BD vs. SZ) and focused binary patient-control contrasts (SZ vs. HC and BD vs. HC). To ensure rigorous evaluation, feature fusion, feature standardisation, Salp Swarm Optimizer (SSO) feature selection, hyperparameter selection, and early stopping are performed exclusively within the training partitions of a participant-level nested cross-validation (five outer and three inner folds), and held-out participants are used only once for evaluation. Evaluated on the UCLA Consortium for Neuropsychiatric Phenomics (CNP) dataset, the framework achieved 79.3% accuracy in four-class transdiagnostic classification. In focused binary tasks, the framework reached 84.3% accuracy for SZ versus HC and 82.7% accuracy for BD versus HC. These results indicate that combining segmentation-guided radiomic descriptors with multi-view deep representations provides a promising research framework for structural MRI analysis of neuropsychiatric disorders, while external validation is required before clinical application.