A Hybrid CNN and ML Framework for Multi-modal Classification of Movement Disorders Using MRI and Brain Structural Features

Atypical Parkinsonian Disorders (APD), also known as Parkinson-plus syndrome, are a group of neurodegenerative diseases that include progressive supranuclear palsy (PSP) and multiple system atrophy (MSA). In the early stages, overlapping clinical features often lead to misdiagnosis as Parkinson’s disease (PD). Identifying reliable imaging biomarkers for early differential diagnosis remains a critical challenge. In this study, we propose a hybrid framework combining convolutional neural networks (CNNs) with machine learning (ML) techniques to classify APD subtypes versus PD and distinguish between the subtypes themselves: PSP vs. PD, MSA vs. PD, and PSP vs. MSA. The model leverages multi-modal input data, including T1-weighted magnetic resonance imaging (MRI), segmentation masks of 12 deep brain structures associated with APD, and their corresponding volumetric measurements. By integrating these complementary modalities, including image data, structural segmentation masks, and quantitative volume features, the hybrid approach achieved promising classification performance with area under the curve (AUC) scores of 0.95 for PSP vs. PD, 0.86 for MSA vs. PD, and 0.92 for PSP vs. MSA. These results highlight the potential of combining spatial and structural information for robust subtype differentiation. In conclusion, this study demonstrates that fusing CNN-based image features with volume-based ML inputs improves classification accuracy for APD subtypes. The proposed approach may contribute to more reliable early-stage diagnosis, facilitating timely and targeted interventions in clinical practice.

💡 Research Summary

This paper addresses the critical need for reliable imaging biomarkers to differentiate atypical Parkinsonian disorders (APD) – specifically progressive supranuclear palsy (PSP) and multiple system atrophy (MSA) – from idiopathic Parkinson’s disease (PD) in the early stages, when clinical symptoms overlap. The authors propose a two‑stage hybrid framework that fuses deep convolutional neural network (CNN)–derived image features with traditional machine‑learning (ML)–based volumetric measurements.

Dataset and preprocessing
A multi‑site cohort of 554 subjects (285 PD, 192 PSP, 77 MSA) provided 3‑D T1‑weighted MRI scans. An automated segmentation pipeline generated binary masks for twelve deep brain structures (midbrain, pons, medulla, superior cerebellar peduncles, lateral ventricles, third/fourth ventricles, bilateral putamen, and caudate nuclei). Volumes of each structure were computed and normalized by intracranial volume (ICV) to reduce inter‑subject variability. Consequently, each participant contributed three modalities: (1) the raw MRI volume, (2) a set of twelve binary masks, and (3) a 12‑dimensional vector of ICV‑corrected volumes.

Model architecture
Stage 1: A 3‑D CNN with a spatial‑pyramid‑inspired design processes the MRI and its corresponding masks. The network contains three parallel branches, each dedicated to one anatomical region (brainstem, ventricles, striatum). Each branch comprises three Conv3D‑BatchNorm‑Pooling blocks followed by global average pooling. The branch‑level feature maps are concatenated and passed through a dense layer with 256 units and dropout, producing a high‑level representation of the image‑mask input.

Stage 2: The 256‑dimensional CNN embedding is concatenated with the 12 volumetric features. This fused vector is fed into a logistic regression classifier with L2 regularization and class‑balanced weighting. The two‑stage pipeline thus integrates voxel‑level texture information with region‑level quantitative data.

Training and evaluation
The dataset was split 80 % for training/validation and 20 % for a held‑out test set. Five‑fold stratified cross‑validation guided hyper‑parameter tuning. The CNN was trained with a batch size of 2, Adam optimizer (learning rate = 1e‑4), weighted binary cross‑entropy loss, early stopping, and learning‑rate reduction on plateau. After training, the dense‑layer activations were extracted, fused with volumetric data, and the logistic regression model was fitted. Performance was measured using accuracy, sensitivity, specificity, Youden’s index, F1‑score, and area under the ROC curve (AUC).

Results
Three binary classification tasks were examined: PSP vs PD, MSA vs PD, and PSP vs MSA. Ablation studies compared single‑modality inputs (volume only, mask only, MRI only) and single‑model approaches (ML only, CNN only). The hybrid model consistently outperformed all baselines.

• PSP vs PD: AUC = 0.95, F1 = 0.91, accuracy ≈ 89.6 % (sensitivity and specificity both > 87 %).
• MSA vs PD: AUC = 0.86, F1 = 0.62, accuracy ≈ 79 %. Performance was lower due to the smaller MSA sample size and class imbalance.
• PSP vs MSA: AUC = 0.92, accuracy ≈ 83 %. The best configuration combined masks and volumes (no raw MRI), highlighting the value of structural guidance.

Interpretability
To address the “black‑box” concern, the authors applied 3‑D Gradient‑Weighted Class Activation Mapping (Grad‑CAM) to each anatomical branch. Class‑specific gradients were pooled, weighted, and passed through a ReLU to generate voxel‑wise attention maps. Population‑averaged maps, up‑sampled to native resolution, consistently highlighted the midbrain, ventricular system, and striatal regions—areas known to be affected in PSP and MSA. This visual evidence demonstrates that the model bases its decisions on biologically plausible pathology rather than spurious image artifacts.

Discussion and limitations
The study’s strengths include (1) multimodal data fusion that leverages complementary information, (2) a clear two‑stage pipeline that remains computationally tractable, (3) rigorous cross‑validation and an external held‑out test set, and (4) built‑in interpretability via 3‑D Grad‑CAM. Limitations comprise the relatively small MSA cohort leading to class imbalance, the high memory demand of 3‑D CNNs, and the use of a simple logistic regression for the final fusion—more sophisticated meta‑learning or ensemble methods might yield further gains. Future work could explore data augmentation, focal loss or SMOTE for imbalance mitigation, end‑to‑end multitask learning (e.g., simultaneous disease classification and severity scoring), and incorporation of additional modalities such as diffusion tensor imaging or functional MRI.

Conclusion
The authors present a novel hybrid framework that integrates deep convolutional features from T1‑weighted MRI and segmentation masks with structure‑specific volumetric measurements. Across three clinically relevant APD classification tasks, the approach achieves AUCs of 0.95 (PSP vs PD), 0.86 (MSA vs PD), and 0.92 (PSP vs MSA), surpassing previously reported results that rely on single‑modality or pure deep‑learning models. Moreover, 3‑D Grad‑CAM visualizations confirm that the classifier focuses on disease‑relevant neuroanatomical regions, enhancing clinical trust. The work demonstrates that multimodal fusion and staged learning can produce both high diagnostic performance and interpretability, positioning the method as a promising decision‑support tool for early differentiation of atypical Parkinsonian disorders.