Can a 3D MRI network spot and forecast Parkinson’s?
A 3D CNN using T1 MRI separated moderate-to-severe Parkinson’s disease from controls at 74% accuracy; transfer learning reached 64% for mild PD, and MRI plus clinical data predicted 2-year progression at >70%.
Source
Multi-Center 3D CNN for Parkinson's disease diagnosis and prognosis using clinical and T1-weighted MRI data
Study at a glance
- Design
- Computational / modelling — Multi-cohort 3D CNN on T1-weighted MRI (± clinical data) with transfer learning and k-means progression clusters
- N
- N=312 · Main: 86 mild PD, 62 moderate-to-severe PD, 60 controls; PPMI 14+14; de novo 38+38 (312 MRI/clinical participants)
- Population
- People with Parkinson’s disease (mild, moderate-to-severe, de novo) and controls from three cohorts including PPMI
- Outcome
- CNN accuracy for PD vs controls and for faster vs slower motor progression over ~2 years
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
Trained a 3D CNN on three cohorts (86 mild, 62 moderate-to-severe PD, 60 controls; PPMI 14+14; de novo 38+38), clustered PD motor progression with k-means on baseline/follow-up UPDRS-III, and tested transfer learning plus MRI+clinical fusion.
What they found
MRI alone: 74% accuracy for moderate-to-severe PD vs controls. Transfer learning improved mild-PD vs control classification to 64%. Progression prediction exceeded 70% with MRI plus clinical features. Activation maps highlighted influential brain regions.
The limits
What it doesn't show
Mid-70s accuracy is not a standalone diagnostic test; UPDRS-based clusters and multi-centre MRI do not replace clinical exam or prove who will need a specific treatment.
Key terms
- 3D CNN
- Three-dimensional convolutional neural network trained on T1-weighted MRI volumes.
- Transfer learning
- Reusing weights from the easier moderate-PD task to improve mild-PD vs control classification.
- UPDRS-III
- Motor exam score used (labeled UDPRS-III in the paper) to k-means cluster progression.
- PPMI
- Parkinson’s Progression Markers Initiative, one of the external mild-PD cohorts.
- Activation maps
- Visualizations of MRI regions that most influenced CNN decisions.
- De novo PD
- Recently diagnosed, largely untreated mild PD (n=38) with matched controls.
Flashcards
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
Moderate-to-severe PD vs control MRI accuracy:
Common questions
Main cohort sizes?
86 mild PD, 62 moderate-to-severe PD, 60 controls.
Moderate PD vs control accuracy?
74% with MRI alone.
Mild PD vs control?
64% after transfer learning.
Progression prediction?
>70% accuracy using MRI plus clinical data.
More on Neurodegeneration