TL;DR
Artificial intelligence (AI) shows high diagnostic performance in detecting Parkinson's disease (PD) using PET imaging, particularly with presynaptic dopamine (DA) tracers for differentiating PD from healthy controls and 18F-FDG PET for distinguishing PD from atypical parkinsonism. This systematic review and meta-analysis highlights AI's potential to improve early and accurate PD diagnosis, especially in resource-limited settings, while also pointing to the need for standardization and rigorous external validation.
What Problem Does This Paper Address?
Parkinson's disease (PD) is a common neurodegenerative disorder, but its early diagnosis is challenging due to overlapping symptoms with other conditions like atypical parkinsonism (AP). Misdiagnosis can lead to incorrect treatment and research outcomes. While PET imaging offers valuable insights into brain function for PD diagnosis, interpreting these complex images often requires specialized expertise, which may not be readily available in all healthcare settings. Artificial intelligence (AI), particularly deep learning (DL) and machine learning (ML), holds promise for automating and improving the accuracy of PD diagnosis from PET scans, but a comprehensive synthesis of its diagnostic performance across various AI approaches and PET tracers is needed.
Methods
This study conducted a systematic review and meta-analysis of published research on AI-assisted PET imaging for PD diagnosis. The researchers systematically searched multiple databases (Ovid MEDLINE, Embase, Web of Science, IEEE Xplore) for relevant studies published up to August 17, 2023. They extracted binary diagnostic accuracy data to calculate key metrics like sensitivity (SE), specificity (SP), and area under the curve (AUC). Meta-analysis was performed to derive pooled estimates of diagnostic performance. Subgroup analyses were conducted to explore the impact of different AI algorithms (DL vs. ML), PET imaging agents (presynaptic DA, postsynaptic DA, 18F-FDG), and sample sizes on diagnostic accuracy. The QUADAS-AI tool was used to assess the risk of bias and applicability of the included studies.
Datasets
The study did not introduce new datasets but rather analyzed data from 23 eligible studies that used various PET imaging datasets. These datasets were sourced from clinical databases of hospitals and research institutions, including the Parkinson's Progression Markers Initiative (PPMI) database. The studies utilized both prospective and retrospective patient data, with some also incorporating data from public databases.
Key Insights and Ideas
- AI's High Diagnostic Potential: AI algorithms, especially when applied to PET imaging, demonstrate a strong capability to accurately differentiate PD from both healthy controls and other parkinsonian syndromes.
- Tracer Specificity: Different PET tracers offer varying strengths. Presynaptic DA PET appears particularly effective for distinguishing PD from normal controls, while 18F-FDG PET excels in differentiating PD from AP.
- DL vs. ML Performance: While both DL and ML show promise, DL algorithms generally exhibited slightly higher diagnostic performance metrics in some comparisons within this meta-analysis.
- Sample Size Matters: Larger sample sizes in AI model training and validation are associated with significantly better diagnostic performance, highlighting the importance of robust datasets.
- Need for Standardization: The heterogeneity observed across studies, stemming from variations in AI methods, data preprocessing, and reporting, underscores a critical need for standardized AI methodologies and reporting practices in medical imaging research.
- External Validation is Crucial: The limited number of studies that performed external validation is a significant limitation, as it impacts the generalizability of AI models to real-world clinical settings.
Key Results
The meta-analysis revealed impressive diagnostic performance metrics:
- PD vs. Normal Control (NC):
- Presynaptic DA PET: Pooled AUC of 0.96 (95% CI: 0.94-0.97).
- 18F-FDG PET: Pooled AUC of 0.90 (95% CI: 0.87-0.93).
- PD vs. Atypical Parkinsonism (AP):
- Presynaptic DA PET: Pooled AUC of 0.93 (95% CI: 0.91-0.95).
- Postsynaptic DA PET: Pooled AUC of 0.79 (95% CI: 0.75-0.82).
- 18F-FDG PET: Pooled AUC of 0.97 (95% CI: 0.96-0.99).
Subgroup analyses showed that DL algorithms generally outperformed ML in some scenarios, and larger sample sizes (≥100) led to higher AUCs (0.94 vs. 0.86 for <100 samples).
Limitations and Future Work
The authors acknowledge several limitations:
- Heterogeneity: Significant heterogeneity was observed across studies, particularly with 18F-FDG PET imaging, which may stem from variations in AI methods, sample sizes, and imaging modalities.
- Data Limitations: Many studies relied on clinical diagnostic criteria without pathological verification, and the composition of AP subtypes varied, potentially affecting results.
- Lack of External Validation: Only a small fraction of studies provided external validation, limiting the generalizability of their findings.
- Reporting Standards: Inconsistent reporting of AI methods and data makes it difficult to compare studies and identify sources of heterogeneity.
- "Black Box" Nature of DL: The lack of transparency in DL models can hinder clinical adoption.
Future work should focus on:
- Standardization: Developing standardized AI methodologies and reporting practices.
- Larger, Diverse Datasets: Creating larger, more diverse, and curated image databases for PD patients.
- Rigorous External Validation: Emphasizing multicenter studies with robust external validation.
- Explainable AI (XAI): Integrating XAI techniques to improve model interpretability and clinician trust.
- Investigating Heterogeneity: Further research to understand and mitigate the sources of heterogeneity.
- Prospective Studies: Conducting more prospective studies to provide stronger evidence.
Summary updated