AI Reads Brain MRI to Diagnose and Stage Parkinson's With High Accuracy
An interpretable machine learning model using MRI brain texture features diagnosed and staged Parkinson's disease with up to 90% accuracy across multiple centers.
Summary
Researchers built an AI model that analyzes MRI scans of six deep brain structures to distinguish healthy individuals from early-stage and mid-to-late-stage Parkinson's disease. Using 200 patients across multiple hospitals, the best-performing model achieved up to 90% accuracy. Crucially, the team used a technique called SHAP analysis to explain which brain regions drove each classification — healthy brains relied on putamen and nucleus accumbens texture, early Parkinson's showed changes in the globus pallidus and putamen, and advanced disease correlated with the substantia nigra and red nucleus. Protein analysis confirmed that these imaging changes map to real biological shifts, including immune, cytoskeletal, and neurodegeneration pathways. This combination of accuracy and explainability could make the tool genuinely useful in clinical practice.
Detailed Summary
Parkinson's disease affects millions worldwide, yet diagnosis remains largely clinical, relying on symptom observation rather than objective biomarkers. Early and accurate staging is critical because treatment strategies differ substantially between early and advanced disease. Objective, imaging-based tools that can explain their reasoning to clinicians are urgently needed.
This multicenter study enrolled 200 patients and divided them into training, internal validation, and external validation cohorts. Six subcortical brain nuclei — caudate, putamen, globus pallidus, red nucleus, substantia nigra, and nucleus accumbens — were precisely segmented from T1-weighted MRI scans. From each nucleus, 107 radiomic texture features were extracted, then narrowed to nine key features through a rigorous three-step selection process. Five machine learning models were compared, with the gradient boosted tree (GBT) model emerging as the top performer.
The GBT model achieved macro-AUC values of 0.890 and 0.876 on internal and external validation sets respectively, representing strong generalizability. SHAP interpretability analysis revealed a biologically coherent staging story: healthy controls were distinguished primarily by putamen and nucleus accumbens texture; early Parkinson's was marked by globus pallidus and putamen structural changes; and mid-to-late disease correlated with substantia nigra and red nucleus features — consistent with known neuropathological progression.
A subset of patients also underwent proteomic analysis. Early Parkinson's versus controls showed 514 differentially expressed proteins enriched in cytoskeletal and immune pathways. Disease progression produced 123 further protein changes tied to proteasome dysfunction and neurodegeneration. Multiple imaging features correlated significantly with specific proteins, linking MRI texture patterns to concrete molecular biology.
The study demonstrates that interpretable AI combining radiomics and proteomics can objectively diagnose and stage Parkinson's disease. Limitations include a relatively small sample size of 200 patients, restriction to T1-weighted MRI, and the proteomic subset being drawn from only one center. The summary is based on the abstract only.
Key Findings
- GBT model achieved macro-AUC of 0.890 internally and 0.876 externally for three-way Parkinson's classification.
- SHAP analysis mapped each disease stage to distinct brain nuclei, mirroring known Parkinson's pathological progression.
- Early Parkinson's was linked to 514 differentially expressed proteins in immune and cytoskeletal pathways.
- MRI texture features correlated significantly with specific proteins such as HGFAC, PLAT, and ASPN, bridging imaging and biology.
- Nine radiomic features — mostly putamen gray-level non-uniformity — were sufficient for accurate diagnosis and staging.
Methodology
The study enrolled 200 patients across multiple centers, split into training (n=76), internal validation (n=33), and external validation (n=91) cohorts. Six subcortical nuclei were segmented from T1-weighted MRI, yielding 107 radiomic features per nucleus, reduced to nine via variance thresholding, univariate selection, and multi-class LASSO. Five machine learning models were evaluated, with SHAP analysis used for interpretability and proteomic correlation performed on a Center 1 subset.
Study Limitations
The total sample of 200 patients is relatively small for a multicenter machine learning study, which may limit statistical power and generalizability. Proteomic analysis was restricted to a subset from a single center, reducing its representativeness. This summary is based on the abstract only, so full methodological details, data quality controls, and extended results are not assessable.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: