Quantitative Susceptibility Mapping Radiomics with Label Noise Compensation for Predicting Deep Brain Stimulation Outcomes in Parkinson’s Disease

Alexandra Roberts, Jinwei Zhang, Ceren Tozlu, Dominick Romano, Sema Akkuş, Heejong Kim, Mert R. Sabuncu, Pascal Spincemaille, Jianqi Li, Yunxin Wang, Xi Wu, Brian H. KopellmedRxiv 2024 · DOI

Abstract

Deep brain stimulation (DBS) is a transformative intervention for Parkinson’s disease (PD) patients experiencing motor complications. Currently, the levodopa challenge test (LCT) serves as the primary clinical predictor for surgical success. However, the LCT is often inconsistent and places a significant burden on patients. This study introduces a novel quantitative susceptibility mapping (QSM) radiomics approach to predict post-surgical outcomes. By extracting features from pre-surgical imaging and implementing a label noise compensation technique, the proposed model achieved a correlation of $r = 0.75$ with actual clinical improvement, significantly outperforming the LCT ($r = 0.12$). These findings suggest that QSM-based radiomics can provide a more accurate, objective, and less invasive method for DBS candidate selection.

Introduction

Deep Brain Stimulation (DBS) of the subthalamic nucleus (STN) is a standard surgical treatment for Parkinson’s disease patients who no longer respond adequately to medication. The selection of surgical candidates traditionally relies on the Levodopa Challenge Test (LCT), which measures a patient's motor improvement in response to levodopa. The underlying assumption is that a positive response to medication predicts a positive response to stimulation.

Despite its widespread use, the LCT is problematic. It requires patients to undergo a difficult "off-medication" period, and its predictive power is notoriously low. In this study, the correlation between LCT results and actual post-surgical Unified Parkinson’s Disease Rating Scale (UPDRS-III) improvement was found to be only $r = 0.12$. There is a critical clinical need for an objective, imaging-based biomarker that can accurately predict surgical outcomes without the need for complex medication manipulation.

Methods

Imaging and Radiomics

The researchers utilized Quantitative Susceptibility Mapping (QSM), an MRI technique that specifically maps iron distribution within the brain. Iron deposition in the substantia nigra (SN) and STN has been previously linked to PD progression.

The study pipeline involved:

  1. Feature Extraction: 1,595 radiomic features—including shape, intensity, and texture—were extracted from the SN and STN using the PyRadiomics library.
  2. Lasso Regression: To manage the high-dimensional feature set relative to the sample size, Lasso regression was employed for automated feature selection and model regularization.

Label Noise Compensation

A primary innovation of this work is the use of "Label Noise Compensation." Clinical scores like the UPDRS-III are inherently subjective and prone to inter-rater variability. By treating these clinical labels as "noisy" and augmenting the training data with Gaussian noise based on label statistics, the model was trained to identify the robust underlying signal rather than overfitting to human measurement errors.

Results

The study analyzed a retrospective cohort of 37 Parkinson’s patients undergoing bilateral STN-DBS. The results demonstrated a stark contrast between the proposed model and current clinical standards:

  • Predictive Accuracy: The QSM radiomics model achieved a correlation of $r = 0.75$ ($p < 0.001$) with actual surgical outcomes, compared to the LCT's $r = 0.12$.
  • Classification Performance: When classifying patients as "responders" or "non-responders," the noise-compensated model reached an AUC of 0.99, a significant improvement over the 0.86 AUC achieved without noise compensation.
  • Biomarker Identification: The model identified specific mathematical textures as key predictors. For instance, higher "coarseness" in the substantia nigra was associated with non-responders, potentially indicating advanced tissue degradation. Conversely, "wavelet entropy" in the STN was a strong predictor for responders, suggesting that specific iron distribution patterns are conducive to successful stimulation.

Discussion

The success of the noise-compensated model suggests that the "ground truth" in clinical neurology is often less precise than the high-resolution imaging data available. By mathematically accounting for this uncertainty, the model uncovered "hidden" biomarkers—such as wavelet-based textures—that are invisible to the naked eye but highly correlated with patient outcomes.

This approach offers several advantages: it is non-invasive, objective, and eliminates the need for patients to undergo the grueling medication withdrawal required by the LCT. Furthermore, the ability to visualize predictive feature maps in deep gray nuclei provides clinicians with a spatial understanding of why a patient may or may not respond to surgery.

Conclusion

QSM-based radiomics, enhanced by label noise compensation, represents a significant advancement in personalized surgical planning for Parkinson’s disease. While the study is limited by a relatively small sample size ($N=37$), the high correlation and classification accuracy warrant further validation in larger, multi-center trials. Future research may extend this framework to other DBS targets and non-motor outcomes, potentially redefining the standard of care for DBS candidate selection.

Summary updated

Abstract

Abstract Parkinson’s disease patients with motor complications are often considered for deep brain stimulation (DBS) surgery. DBS candidate selection involves an assessment known as the levodopa challenge test (LCT). The LCT aims to predict DBS outcomes by measuring symptom improvement accompanying changes in levodopa dosage. While used in the patient selection process, inconsistent LCT predictions have been widely documented, verified here with Pearson’s correlation r = 0.12. Estimating symptom improvement to separate DBS responders and non-responders remains an unmet need. Quantitative susceptibility mapping (QSM) is routinely acquired for pre-surgical planning and depicts the iron distribution in substantia nigra and subthalamic nuclei. Iron deposition in these nuclei has correlated with disease progression and motor symptom severity. A novel QSM radiomic approach is presented using a regression model and features extracted from the pre-surgical targeting acquisition. Noise compensation in training labels improves outcome prediction in regression and classification models. The model predicts improvement in the unified Parkinson’s disease rating scale (UPDRS-III) ( r = 0.75). Predictive feature maps in deep gray nuclei offer contrast between responders and non-responders. The QSM radiomic approach has potential to improve DBS candidate selection by accurately estimating symptom improvement, eliminating difficult medication manipulation, and avoiding time-consuming evaluations to reduce patient and clinician burden.
BibTeX
@misc{roberts2024quantitative,
  author = {Alexandra Roberts and Jinwei Zhang and Ceren Tozlu and Dominick Romano and Sema Akkuş and Heejong Kim and Mert R. Sabuncu and Pascal Spincemaille and Jianqi Li and Yunxin Wang and Xi Wu and Brian H. Kopell},
  title = {{Quantitative Susceptibility Mapping Radiomics with Label Noise Compensation for Predicting Deep Brain Stimulation Outcomes in Parkinson’s Disease}},
  year = {2024},
  doi = {10.1101/2024.12.26.24319663},
  howpublished = {medRxiv},
}
Figure