Dopamine Transporter SPECT Image Classification for Neurodegenerative Parkinsonism via Diffusion Maps and Machine Learning Classifiers

Jun-En Ding, Chi‐Hsiang Chu, Mong‐Na Lo Huang, Chien-Ching Hsu

Abstract

Neurodegenerative parkinsonism can be assessed by dopamine transporter single photon emission computed tomography (DaT-SPECT). Although generating images is time consuming, these images can show interobserver variability and they have been visually interpreted by nuclear medicine physicians to date. Accordingly, this study aims to provide an automatic and robust method based on Diffusion Maps and machine learning classifiers to classify the SPECT images into two types, namely Normal and Abnormal DaT-SPECT image groups. In the proposed method, the 3D images of N patients are mapped to an N by N pairwise distance matrix and are visualized in Diffusion Maps coordinates. The images of the training set are embedded into a low-dimensional space by using diffusion maps. Moreover, we use Nyström's out-of-sample extension, which embeds new sample points as the testing set in the reduced space. Testing samples in the embedded space are then classified into two types through the ensemble classifier with Linear Discriminant Analysis (LDA) and voting procedure through twenty-five-fold cross-validation results. The feasibility of the method is demonstrated via Parkinsonism Progression Markers Initiative (PPMI) dataset of 1097 subjects and a clinical cohort from Kaohsiung Chang Gung Memorial Hospital (KCGMH-TW) of 630 patients. We compare performances using Diffusion Maps with those of three alternative manifold methods for dimension reduction, namely Locally Linear Embedding (LLE), Isomorphic Mapping Algorithm (Isomap), and Kernel Principal Component Analysis (Kernel PCA). We also compare results using 2D and 3D CNN methods. The diffusion maps method has an average accuracy of 98% for the PPMI and 90% for the KCGMH-TW dataset with twenty-five fold cross-validation results. It outperforms the other three methods concerning the overall accuracy and the robustness in the training and testing samples.

Introduction

Diagnosing Parkinson’s disease (PD) and related neurodegenerative disorders often relies on Dopamine Transporter SPECT (DaT-SPECT) imaging. These scans visualize the density of dopamine transporters in the striatum. In a healthy brain, these transporters appear as bright, "comma-shaped" regions. In contrast, neurodegeneration typically results in a loss of this intensity, causing the regions to appear circular, faint, or asymmetric.

Visual difference between normal and abnormal DaT-SPECT scans showing the characteristic 'comma' vs. 'circular' patterns.

Currently, these images are manually interpreted by nuclear medicine physicians, a process that is time-consuming and prone to inter-observer variability. While deep learning models like Convolutional Neural Networks (CNNs) offer automation, they often act as "black boxes," requiring massive datasets and significant computational power. This study proposes a more robust and explainable alternative using manifold learning.

Methodology: Diffusion Maps and Manifold Learning

The core of the proposed method is Diffusion Maps (DM), a manifold learning technique designed to discover the intrinsic geometric structure of high-dimensional data.

  1. Dimensionality Reduction: Instead of processing raw pixels directly, the system calculates a pairwise distance matrix between the 3D images of all patients. DM then "unfolds" this complex data into a low-dimensional space (e.g., 200 dimensions), capturing the non-linear relationships between samples.
  2. Nyström’s Out-of-Sample Extension: To handle new patients without re-calculating the entire manifold, the authors utilized Nyström’s extension. This allows new "testing" data to be projected into the established "training" space efficiently.
  3. Ensemble Classification: Once embedded in the low-dimensional space, a Linear Discriminant Analysis (LDA) classifier is used. To ensure stability, the researchers employed a 25-fold cross-validation voting procedure, where the final diagnosis is determined by the majority vote across multiple data splits.
Conceptual mapping of high-dimensional patient image data into a low-dimensional space using Diffusion Maps.

Experimental Setup and Datasets

The method was validated using two distinct datasets to test both benchmark performance and real-world clinical utility:

  • PPMI (Parkinsonism Progression Markers Initiative): A standardized research dataset containing 1,097 subjects (876 PD patients and 414 healthy controls).
  • KCGMH-TW: A clinical cohort from Kaohsiung Chang Gung Memorial Hospital in Taiwan, consisting of 630 patients (353 normal, 377 abnormal). This dataset represents noisier, "real-world" clinical conditions.

Key Results and Discussion

The Diffusion Maps approach demonstrated significant advantages over both traditional manifold methods and modern deep learning architectures:

  • Superior Accuracy: The DM+LDA method achieved 98% accuracy on the PPMI dataset, outperforming 2D-CNNs like VGG-16 (93%) and AlexNet (88%). On the noisier KCGMH-TW clinical dataset, it maintained a high 90% accuracy when using three consecutive brain slices, surpassing 3D-CNNs (82%).
  • Stability and Efficiency: Unlike deep learning models, which often show high variability in performance, the DM approach was remarkably stable (standard deviation of 0.02 to 0.05). Furthermore, it requires significantly less training time and hardware, making it more accessible for clinical environments.
  • Visual Explainability: In the 3D diffusion space, patients naturally align along a "U-shaped" trajectory. Healthy controls cluster at one end, while severe cases move toward the other. This provides clinicians with a visual "pathway" of disease progression, offering transparency that "black-box" models lack.

Conclusion

By combining Diffusion Maps with simple linear classifiers, this study provides a robust, efficient, and explainable tool for DaT-SPECT classification. The method not only matches or exceeds the accuracy of complex deep learning models but also provides a stable framework for clinical decision-making. Future work may expand this binary classification to distinguish between different types of parkinsonism and incorporate multi-modal data such as patient age and clinical symptoms.

Summary updated

Abstract

Neurodegenerative parkinsonism can be assessed by dopamine transporter single photon emission computed tomography (DaT-SPECT). Although generating images is time consuming, these images can show interobserver variability and they have been visually interpreted by nuclear medicine physicians to date. Accordingly, this study aims to provide an automatic and robust method based on Diffusion Maps and machine learning classifiers to classify the SPECT images into two types, namely Normal and Abnormal DaT-SPECT image groups. In the proposed method, the 3D images of N patients are mapped to an N by N pairwise distance matrix and are visualized in Diffusion Maps coordinates. The images of the training set are embedded into a low-dimensional space by using diffusion maps. Moreover, we use Nyström's out-of-sample extension, which embeds new sample points as the testing set in the reduced space. Testing samples in the embedded space are then classified into two types through the ensemble classifier with Linear Discriminant Analysis (LDA) and voting procedure through twenty-five-fold cross-validation results. The feasibility of the method is demonstrated via Parkinsonism Progression Markers Initiative (PPMI) dataset of 1097 subjects and a clinical cohort from Kaohsiung Chang Gung Memorial Hospital (KCGMH-TW) of 630 patients. We compare performances using Diffusion Maps with those of three alternative manifold methods for dimension reduction, namely Locally Linear Embedding (LLE), Isomorphic Mapping Algorithm (Isomap), and Kernel Principal Component Analysis (Kernel PCA). We also compare results using 2D and 3D CNN methods. The diffusion maps method has an average accuracy of 98% for the PPMI and 90% for the KCGMH-TW dataset with twenty-five fold cross-validation results. It outperforms the other three methods concerning the overall accuracy and the robustness in the training and testing samples.
BibTeX
@misc{ding2021dopamine,
  author = {Jun-En Ding and Chi‐Hsiang Chu and Mong‐Na Lo Huang and Chien-Ching Hsu},
  title = {{Dopamine Transporter SPECT Image Classification for Neurodegenerative Parkinsonism via Diffusion Maps and Machine Learning Classifiers}},
  year = {2021},
  eprint = {2104.02066},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.02066},
}
Figure