Abstract
This study suggested the developed deep neural network can decode in-depth information from DAT and showed potential to assist the differential diagnosis of parkinsonism. The functional regions supporting the diagnosis decision were generally consistent with known parkinsonian pathology but provided more specific guidance for feature selection and quantitative analysis.
Introduction
Distinguishing idiopathic Parkinson’s disease (IPD) from atypical parkinsonian syndromes (APS)—such as Multiple System Atrophy (MSA) and Progressive Supranuclear Palsy (PSP)—is a significant clinical challenge, particularly in the early stages where symptoms overlap. Traditional clinical analysis of Dopamine Transporter (DAT) PET imaging relies on the "Binding Ratio" (BR), a simple average of tracer uptake in the striatum. However, this method often fails to capture the complex spatial distribution of dopamine loss. To overcome these limitations, researchers developed DAT-Net, a 3D deep residual convolutional neural network designed to extract high-dimensional diagnostic features from [11C]CFT PET scans.
Methodology and Dataset
The study utilized the Huashan Parkinsonian PET Imaging (HPPI) Database, comprising 1,017 subjects, making it one of the largest datasets of its kind. The cohort included healthy controls and patients with IPD, MSA, and PSP.
The proposed framework follows a three-step process:
- AI-based Diagnosis: A 3D Deep Residual CNN (DAT-Net) processes entire PET volumes to classify patients.
- Model Interpretation: Full-gradient saliency maps identify the voxels most critical to the AI's decision-making process.
- Derived Metrics: "Deep-Learning-guided Binding Ratios" (DL-BR) are calculated based on these salient regions, providing a more sensitive alternative to traditional ROI-based metrics.
Results and Key Insights
DAT-Net demonstrated superior diagnostic accuracy compared to traditional methods. In cross-validation, the model achieved AUCs of 0.953 (IPD), 0.948 (MSA), and 0.900 (PSP). These results were maintained in an independent blind-test cohort, with sensitivities reaching 90.7% for IPD.
A critical insight gained from the saliency maps was the importance of the midbrain in differential diagnosis. While human experts focus on the striatum, the AI identified midbrain uptake patterns as highly discriminatory. This aligns with the known pathology of [11C]CFT, which binds to serotonin transporters in the midbrain that are differentially affected in MSA and PSP. Furthermore, the DL-BR metrics showed highly significant differences between disease groups ($P < 0.001$) where traditional ratios often showed overlap.
Discussion and Future Directions
The success of DAT-Net suggests that deep learning can "decode" subtle pathological signatures in neuroimaging that are invisible to the human eye. The model's robustness across varying disease durations makes it a promising tool for early-stage diagnosis. Future research will focus on prospective validation and adapting the model for use with [123I]FP-CIT SPECT scans, which are more widely available in clinical practice.
Summary updated