Computer-Aided Classification Framework of Parkinsonian Disorders Using 11C-CFT PET Imaging

Jiahang Xu, Qian Xu, Shihong Liu, Ling Li, Lei Li, Tzu‐Chen Yen, Jian Wu, Jian Wang, Chuantao Zuo, Ping Wu, Xiahai ZhuangFrontiers in Aging Neuroscience 2022 · DOI

TL;DR

This paper introduces a novel computer-aided classification framework using 11C-CFT PET imaging to differentiate between Parkinson's disease (PD), multiple system atrophy (MSA), and progressive supranuclear palsy (PSP). The framework achieves promising diagnostic accuracy, highlighting the potential of AI-driven PET analysis for improving the differential diagnosis of these complex neurological disorders.

What Problem Does This Paper Address?

Parkinsonian disorders (PDs) are a group of neurological conditions that share similar symptoms like tremor, rigidity, and bradykinesia. Differentiating between Parkinson's disease (PD) and other atypical parkinsonian syndromes (APSs) such as multiple system atrophy (MSA) and progressive supranuclear palsy (PSP) is crucial for effective treatment, patient management, and clinical trial design. However, this differential diagnosis is challenging due to overlapping symptoms, especially in the early stages of the disease. Current diagnostic methods often rely on a combination of clinical assessment and imaging techniques, which can be costly and time-consuming. While dopamine transporter (DAT) PET imaging is useful for distinguishing PD from non-dopaminergic conditions, its utility in differentiating between different PD subtypes has been limited. This paper aims to address this gap by developing and evaluating a computer-aided classification framework using 11C-CFT PET imaging to improve the accuracy of differentiating between PD, MSA, and PSP.

Methods

The researchers developed a novel, multi-step computer-aided classification framework. The core of this framework involves:

  1. MRI-Assisted PET Segmentation: Magnetic Resonance Imaging (MRI) was used to guide the segmentation of key brain regions in the 11C-CFT PET images. Specifically, a U-Net architecture, a type of convolutional neural network (CNN) known for its effectiveness in image segmentation, was employed to segment the striatum (including caudate, putamen, pallidus, and midbrain). Bilateral occipital areas were also segmented. The MRI segmentation masks were then aligned with the PET images.
  2. Feature Extraction and Prediction: From the segmented PET images, features were extracted. These features represented the volume of tracer uptake in various striatal subregions. To normalize these values, the striatal-to-occipital ratio (SOR) was calculated. The striatum was further divided into anterior, middle, and posterior subregions using k-means clustering.
  3. Automatic Subject Classification: A Support Vector Machine (SVM) classifier was used for the final classification of subjects into PD, MSA, or PSP categories. The SVM was chosen for its ability to handle high-dimensional feature spaces efficiently.

The study also explored different training strategies, including combinations of early-stage (ES) and advanced-stage (AS) patients, to optimize classification performance.

Datasets

The study utilized data from:

  • 107 patients with PDs: This cohort included individuals clinically diagnosed with Parkinson's disease (PD), multiple system atrophy (MSA), and progressive supranuclear palsy (PSP).
  • 22 normal controls (NC): This group served as a reference for DAT binding levels.

The patients were further categorized into early-stage (ES, disease duration < 24 months) and advanced-stage (AS, disease duration ≥ 24 months) groups.

Key Insights and Ideas

  • Striatal Subregions are Key: The study strongly suggests that the caudate and putamen are the most diagnostically relevant regions for differentiating between PD, MSA, and PSP using 11C-CFT PET imaging. This aligns with the known dopaminergic dysfunction in these areas in these disorders.
  • Midbrain's Limited Role: Interestingly, the contribution of the midbrain to the classification accuracy was found to be negligible. This suggests that while midbrain pathology is central to PD, its imaging signature in 11C-CFT PET might not be as discriminative for differentiating these specific PD subtypes compared to striatal changes.
  • Early vs. Advanced Stage Training: Combining data from both early and advanced disease stages for training the classification model generally yielded better performance across most metrics (sensitivity, specificity, PPV, NPV) compared to training solely on one stage, especially for diagnosing PD and MSA. This highlights the benefit of a comprehensive dataset for robust model training.
  • Asymmetric Putamen as a Discriminator: Asymmetric indexes of the anterior and middle putamen showed significant elevation in the PD group compared to other groups, suggesting this specific feature could be a valuable indicator.

Key Results

  • Binary Classification Accuracy: The framework achieved high accuracy in binary classification: 85.0% for PD vs. others, 82.2% for MSA vs. others, and 89.7% for PSP vs. others.
  • Multiclass Classification Accuracy: The overall accuracy for classifying subjects into PD, MSA, or PSP was 80.4%.
  • Regional Contribution: The caudate contributed 49.2%, the putamen 41.6%, and the pallidum 9.2% to the classification. Specifically, anterior caudate, middle putamen, and middle caudate were the most influential subregions.
  • Training Strategy Performance: The strategy combining both early and advanced stage patients for training (Strategy 1) resulted in the highest accuracy (77.8%) for early-stage patients.

Limitations and Future Work

The authors acknowledge several limitations:

  • Sample Size: The study's sample size, while substantial, is noted as relatively small. They emphasize the need for validation studies on larger, multi-center datasets to ensure external validity.
  • Pipeline Transparency: The use of deep neural networks for feature extraction might compromise the transparency of the entire pipeline, making it a "black box" to some extent.
  • User-Friendliness: The proposed framework, while effective, could be further optimized for user-friendliness and improved classification performance.

Future work should focus on independent validation of the framework in diverse patient populations and potentially exploring more interpretable AI models to enhance understanding of the diagnostic features. The authors believe their approach holds promise for improving differential diagnosis in both research and clinical settings.

Summary updated

Abstract

The proposed computer-aided classification framework based on <sup>11</sup>C-CFT PET imaging holds promise for improving the differential diagnosis of PDs.
BibTeX
@inproceedings{xu2022computer,
  author = {Jiahang Xu and Qian Xu and Shihong Liu and Ling Li and Lei Li and Tzu‐Chen Yen and Jian Wu and Jian Wang and Chuantao Zuo and Ping Wu and Xiahai Zhuang},
  title = {{Computer-Aided Classification Framework of Parkinsonian Disorders Using 11C-CFT PET Imaging}},
  booktitle = {Frontiers in Aging Neuroscience},
  year = {2022},
  doi = {10.3389/fnagi.2021.792951},
}
Figure