Unlocking the potential: T1-weighed MRI as a powerful predictor of levodopa response in Parkinson’s disease

Junyi Yan, Xufang Luo, Jiahang Xu, Dongsheng Li, Lili Qiu, Dianyou Li, Peng Cao, Chencheng ZhangInsights 2024 · DOI

TL;DR

This study demonstrates that T1-weighted MRI features, extracted using a deep learning model called MedicalNet, can effectively predict a patient's response to levodopa treatment for Parkinson's disease. These imaging-based predictions are more robust and generalizable than traditional clinical information, offering a promising avenue for personalized treatment strategies.

What Problem Does This Paper Address?

Parkinson's disease (PD) is a progressive neurodegenerative disorder that significantly impacts quality of life. Levodopa is the cornerstone treatment, but predicting how well a patient will respond is crucial for effective management, especially when considering interventions like deep brain stimulation. Currently, the standard method for assessing levodopa response is the levodopa challenge test (LCT), which is time-consuming and lacks predictive modeling. While T1-weighted MRI is a widely used imaging technique in PD for diagnosis and differentiation, its potential for predicting levodopa response has been underexplored. Previous studies have shown promise but often suffer from small sample sizes, limited validation, or a failure to isolate the predictive power of imaging features. This paper aims to fill this gap by rigorously evaluating T1-weighted MRI's predictive capability for levodopa response, comparing it against clinical data and exploring the underlying brain regions involved.

Methods

The researchers employed a multi-stage approach involving feature extraction, feature selection, and machine learning model training and validation.

Feature Extraction: Four distinct methods were used to extract features from T1-weighted MRI scans:

  1. Age-corrected regional gray matter intensity: This method, based on prior work, extracts intensity from discriminative voxels and then uses Principal Component Analysis (PCA) to reduce dimensionality.
  2. Subcortical ROI textures: Features were extracted from 16 subcortical regions of interest (ROIs) using PyRadiomics, focusing on texture characteristics.
  3. Morphological graph network: This approach constructs a graph representing individual brain networks and extracts graph metrics.
  4. MedicalNet: This is the novel approach proposed by the authors. They adapted a pre-trained ResNet-based deep learning model (MedicalNet), originally designed for medical images, by replacing its segmentation layers with max-pooling and flattening layers. This modified model acts as a pure feature extractor.

Feature Selection: To refine the extracted features and mitigate redundancy, a sequential feature selection process was applied using Minimum Redundancy Maximum Relevance (mRMR), Least Absolute Shrinkage and Selection Operator (LASSO), and Recursive Feature Elimination (RFE).

Machine Learning Models: Three common machine learning models were trained and tested: Support Vector Machine (SVM), XGBoost, and Multi-layer Perceptron (MLP).

Validation: The models were trained and tested on the Parkinson's Progression Markers Initiative (PPMI) dataset. Crucially, the performance was then validated on an independent external clinical dataset from Ruijin Hospital to assess generalizability.

Visualization: Grad-CAM was used to visualize the important brain regions identified by the MedicalNet extractor.

Datasets

  • Parkinson's Progression Markers Initiative (PPMI) dataset: This was the primary dataset used for training and testing. It included 219 PD patient records and 193 healthy control records (used for age correction). The PD records were split into an 8:2 ratio for training (173 records) and testing (46 records).
  • External Clinical Dataset (Ruijin Hospital): This dataset, comprising 217 PD patient records, was used for external validation of the trained models. This dataset had some potential biases, including a higher proportion of "good" responders and longer disease duration.

Key Insights and Ideas

  • Deep Learning for MRI Feature Extraction is Superior: The MedicalNet-based feature extractor significantly outperformed traditional methods (gray matter intensity, subcortical textures, morphological graphs) in predicting levodopa response. This highlights the power of deep learning for extracting subtle, predictive patterns from T1-weighted MRI.
  • Imaging Features Trump Clinical Data: T1-weighted MRI features, especially those from MedicalNet, proved to be more robust and generalizable predictors of levodopa response than demographic and clinical information alone. This suggests that objective imaging biomarkers can offer a more reliable assessment.
  • Union of Features is Best: While imaging features alone were powerful, combining them with clinical information (the "union set") generally yielded the best predictive performance across most models. This indicates that both types of information are valuable, but imaging provides a stronger foundation.
  • Key Brain Regions Identified: The visualization of MedicalNet features pointed to specific brain regions, including the superior temporal gyrus, cingulate gyrus, thalamus, putamen, and hippocampus, as being particularly important for predicting levodopa response. This offers potential insights into the neurobiological underpinnings of levodopa response.

Key Results

  • MedicalNet Achieves Highest AUC: On the PPMI test set, the union of MedicalNet features with clinical data achieved the highest AUC of 0.83 ± 0.01 using SVM. This was significantly better than clinical features alone (AUC 0.73 ± 0.01) and other imaging feature sets.
  • Consistent Improvement: MedicalNet features consistently improved performance across all three machine learning models (SVM, XGBoost, MLP) compared to clinical features alone.
  • External Validation Success: On the external clinical dataset, the MedicalNet-extracted features, when combined with clinical data (Union set), achieved an AUC of 0.59 ± 0.005 using SVM. While lower than on the PPMI test set, this still demonstrated a statistically significant improvement over clinical data alone (AUC 0.53 ± 0.01), underscoring the robustness of the imaging features.
  • Subcortical Texture Also Shows Promise: Subcortical texture features also showed a statistically significant improvement when combined with clinical data on the external set (AUC 0.57 ± 0.005 vs. 0.53 ± 0.01).

Limitations and Future Work

  • Sample Size Limitations: While the study used a substantial number of records, the authors acknowledge that the sample size was still limited, potentially leading to biased models and hindering the training of deep learning models for highly specific tasks.
  • External Validation Scope: The use of only one retrospective external clinical set limits the ability to fully assess the generalizability of the predictive features.
  • Interpretability: While key brain regions were identified, the authors suggest the need for more interpretable models to fully elucidate the pathological roles of these regions in levodopa response.
  • Future Directions: The authors propose exploring other MRI sequences (e.g., T2* or quantitative susceptibility mapping) and integrating information from newly identified brain regions to develop even more robust and generalizable models for levodopa response prediction. They also emphasize the need for future prediction models to rely more heavily on imaging features, given the potential biases in clinical information.

Summary updated

Abstract

This study investigated the predictive value of T1w features for levodopa response. MedicalNet extractor outperformed all other previously published methods with key region visualization. T1w features are more effective than clinical information in levodopa response prediction.
BibTeX
@inproceedings{yan2024unlocking,
  author = {Junyi Yan and Xufang Luo and Jiahang Xu and Dongsheng Li and Lili Qiu and Dianyou Li and Peng Cao and Chencheng Zhang},
  title = {{Unlocking the potential: T1-weighed MRI as a powerful predictor of levodopa response in Parkinson’s disease}},
  booktitle = {Insights},
  year = {2024},
  doi = {10.1186/s13244-024-01690-z},
}
Figure