MORPH2DIAG: Automated Structural MRI Preprocessing and Tissue Segmentation for Interpretable Machine and Deep Learning-Based Neuroanatomical Classification

Abstract Structural MRI provides a noninvasive window into brain morphology, yet the reproducibility and interpretability of morphometric analyses remain limited by inconsistent preprocessing, variable spatial alignment, and heterogeneous feature construction. We introduce MORPH2DIAG, a fully automated, atlas-free morphometric pipeline that integrates standardized preprocessing, tissue segmentation, spatial normalization, data-driven subtyping, and machine- and deep-learning classification within a single, modular framework. The pipeline performs intensity normalization, morphological cleanup, PCA-informed affine alignment, isotropic rescaling, and Gaussian Mixture Model (GMM) segmentation to generate quantitative gray-matter (GM), white-matter (WM), and cerebrospinal-fluid (CSF) maps. Global tissue fractions were used to derive latent neuroanatomical subtypes via unsupervised K-means clustering, revealing progressive GM–CSF gradients consistent with patterns commonly observed along normative-to-atrophic structural continua observed in neurodegeneration. To capture finer-grained spatial heterogeneity, a voxel-wise K-means parcellation yielded parcel-level intensity means and variances that served as regional morphometric descriptors. These global and parcel-level features were integrated into a unified evaluation suite comparing classical machine learning models (Random Forests, Logistic Regression, XGBoost) with lean, deep, and hybrid multilayer perceptrons (MLPs) trained using focal loss, label smoothing, stochastic weight averaging, and nested cross-validation with PCA-based dimensionality reduction. Across methods, the hybrid MLP achieved the highest macro-F1 and balanced accuracy, demonstrating strong discriminative performance for the discovered morphometric subtypes. Collectively, MORPH2DIAG establishes a fully automated, atlas-free framework that unites unsupervised structural subtype discovery with interpretable machine and deep learning, providing a reproducible foundation for MRI-based morphometric profiling and automated detection of neurodegenerative-like patterns. Competing Interest Statement The authors have declared no competing interest. Footnotes shreya_bangera{at}berkeley.edu, pospisil.libor{at}berkeley.edu, aiktocke{at}berkeley.edu