Parkinson’s disease (PD) is the second most prevalent neurodegenerative disorder worldwide, yet early diagnosis remains a significant challenge. Standard clinical assessments and conventional MRI lack specificity, leading to notable misclassification rates, especially in early disease stages. While advanced imaging modalities can improve diagnostic precision, their cost and complexity limit widespread use. A recent study explored the feasibility of using machine learning (ML) models trained on radiomics features from routine T2-weighted fluid-attenuated inversion recovery (T2W FLAIR) images to distinguish PD patients from healthy controls (HCs), potentially offering a cost-effective, accessible screening tool.
Radiomics and Machine Learning Integration
Radiomics enables the extraction of high-dimensional features from medical images, capturing patterns that may not be visually discernible. This study used T2W FLAIR images from a large multicentre cohort comprising 1727 participants, including both PD patients and HCs, to extract radiomics features from four brain regions known to be implicated in PD: the substantia nigra (SN), red nucleus (RN), globus pallidus (GP) and putamen (PU). After manual segmentation and validation of these regions of interest, a total of 7124 features were extracted and refined using dimension reduction techniques such as LASSO and mRMR, ultimately selecting 20 highly diagnostic features.
Must Read: Enhancing Brain MRI Imaging with Deep Learning
These features included five from SN, two from RN, three from GP and ten from PU. All were higher-order texture features, suggesting that PD-related changes are too subtle to detect using basic intensity or shape-based analyses. The SN-derived features showed the highest diagnostic importance, reflecting the critical role of dopaminergic neuron degeneration in PD. PU features also contributed significantly, aligning with known patterns of dopaminergic loss and structural changes in the striatum during disease progression. Together, these features formed the basis for model training and testing using six ML algorithms.
Model Performance and Validation
The dataset was split into a training cohort and two test cohorts (internal and external) to assess model robustness and generalisability. The six ML algorithms used were support vector machine (SVM), multilayer perceptron (MLP), K-nearest neighbour (KNN), random forest (RF), Gaussian Naive Bayes (GNB) and adaptive boosting (AB). All were trained using the selected radiomics features from the internal training set and tested on both internal and external datasets.
Performance was evaluated using standard classification metrics, including area under the curve (AUC), accuracy, sensitivity, specificity, precision and F1 score. In the internal test cohort, all six models achieved high AUC values (0.96–0.98), with SVM and MLP achieving the highest accuracy of 0.90. External test results were slightly lower but still promising: the MLP model led with an AUC of 0.85 and an accuracy of 0.78, closely followed by SVM and KNN. These results confirmed the feasibility of using conventional MRI enhanced by ML for PD screening, with three models maintaining strong performance across different institutions and imaging platforms.
Clinical Relevance and Future Implications
The study highlights a significant advance in PD diagnostics by leveraging routinely performed MRI scans and ML. Importantly, it demonstrates that non-invasive, cost-effective imaging sequences like T2W FLAIR can yield clinically useful information when analysed with computational tools. The segmentation of relevant brain regions and the extraction of detailed features were reproducible, as shown by high Dice similarity coefficients between neuroradiologists. Moreover, the use of data from multiple centres and scanner types enhances confidence in the models' generalisability.
Although the approach is semi-automated and relies on expert manual segmentation, the growing development of automated deep learning-based segmentation tools holds promise for streamlining future diagnostic workflows. Furthermore, while the current model focuses solely on radiomic data, future integration with clinical markers could further enhance diagnostic precision. The study also confirms that subtle structural and texture differences in the SN, RN, GP and PU—key regions in PD pathophysiology—can be captured using conventional imaging, thus expanding the potential of these routinely acquired scans.
The study demonstrated that machine learning models trained on radiomics features from T2W FLAIR images can reliably distinguish Parkinson’s disease patients from healthy controls. The findings support the integration of computational analysis into standard imaging workflows, offering an accessible and scalable solution for early PD screening. Continued research, including prospective validation and automated segmentation, will be crucial in translating these findings into clinical practice.
Source: Insights into Imaging
Image Credit: iStock
