Source-linked AI summary
An Explainable Machine Learning Model for Early Detection of Parkinson's Disease using LIME on DaTscan Imagery
Pavan Rajkumar Magesh, Richard Delwin Myloth, Rijo Jackson Tom
TL;DR
Parkinson’s disease diagnosis from SPECT DaTscans needs accurate classification and interpretable reasoning for medical experts. The paper combines VGG16 transfer learning with LIME visual explanations, achieving 95.2% accuracy and proposing support for earlier diagnosis. Its scope is bounded by questionable labels, single-slice inputs, and the need for clinical validation.
Problem
Manual SPECT interpretation can involve human error, motivating an interpretable system that classifies Parkinson’s disease from DaTscans and explains its predictions.
Method
The study uses a VGG16 convolutional neural network with transfer learning and LIME superpixel explanations for DaTscan classification.
Results
95.2% accuracy was achieved for classifying SPECT DaTscan images as Parkinson’s disease or not.
Takeaways & Limitations
The proposed system may aid early Parkinson’s diagnosis by making predictions faster, more intuitive, and visually interpretable for healthcare workers.
Takeaways & Limitations
The data labels may be questionable, the model uses single-slice inputs, and clinical validation with novel real-time data remains necessary.
Abstract
from arXiv · showhide
Parkinson's disease (PD) is a degenerative and progressive neurological condition. Early diagnosis can improve treatment for patients and is performed through dopaminergic imaging techniques like the SPECT DaTscan. In this study, we propose a machine learning model that accurately classifies any given DaTscan as having Parkinson's disease or not, in addition to providing a plausible reason for the prediction. This is kind of reasoning is done through the use of visual indicators generated using Local Interpretable Model-Agnostic Explainer (LIME) methods. DaTscans were drawn from the Parkinson's Progression Markers Initiative database and trained on a CNN (VGG16) using transfer learning, yielding an accuracy of 95.2%, a sensitivity of 97.5%, and a specificity of 90.9%. Keeping model interpretability of paramount importance, especially in the healthcare field, this study utilises LIME explanations to distinguish PD from non-PD, using visual superpixels on the DaTscans. It could be concluded that the proposed system, in union with its measured interpretability and accuracy may effectively aid medical workers in the early diagnosis of Parkinson's Disease.
1. Introduction
Parkinson’s disease affects dopaminergic neurons and can be difficult to diagnose manually from medical images. This study proposes deep-learning classification of SPECT DaTscans with LIME-based visual explanations for medical experts.
- Parkinson’s disease degenerates dopamine-producing neurons in the substantia nigra and can cause motor and non-motor symptoms.
- Manual SPECT interpretation and symptom-based treatment can delay diagnosis or expose results to human error.Expert diagnosis validity rose from 79.6% initially to 83.9% after follow-up checks using DaTscans.
- The proposed system classifies DaTscans as PD or non-PD by detecting patterns around the putamen and caudate regions.
- The paper develops an accurate deep-learning model, analyzes VGG16 performance, and adds LIME explanations to the classification task.
- The experiment covers dataset preparation, preprocessing, splitting, neural-network architecture, transfer learning, results, and LIME-based explainability.
2. Related Work
Earlier Parkinson’s classification studies used handcrafted or voxel-based features with classical classifiers, while newer work applied neural networks to DaTscan patterns. The paper positions itself as an explainable Parkinson’s model addressing a gap in healthcare interpretability.
- Early DaTscan studies used Naive Bayes with PCA or SVMs with voxel, shape-surface, and striatal binding-ratio features.
- CNN-based approaches were later proposed to detect Parkinson’s-associated patterns in DaTscan images.
- Prior healthcare studies often prioritized accuracy, while interpretability frameworks were developed for medical image classification.
- The paper states that interpretable models existed for other neurodegenerative diseases but not for Parkinson’s disease.
3. Early Parkinson’s Disease Detection CNN Model
The study develops a VGG16-based CNN pipeline for classifying PD from PPMI DaTscan images and evaluates threshold-dependent diagnostic performance. Using an optimized threshold, the model achieved improved accuracy, specificity, precision, ROC area, Cohen’s kappa, and F1 score.
- Dataset: 642 DaTscan SPECT images from the PPMI database comprised 430 PD and 212 non-PD cases, using only initial screening scans.The dataset excluded follow-up scans to preserve patient uniqueness and reduce potential overfitting from within-patient similarity.
- Image preprocessing: The 41st slice was extracted because it most prominently depicted the putamen and caudate regions of interest, then cropped and converted to JPEG.Cropping removed black corners and standardized image size across scans.
- Dataset splitting: The dataset was divided into training, validation, and test sets using an 80:10:10 ratio.Each split was further divided into healthy-control and PD categories.
- Model architecture and training: Transfer learning was used to initialize the CNN from a model pretrained on ImageNet rather than training the complete network from random values.The approach addresses the difficulty of training a full CNN with limited medical-image data and computational resources.
- Threshold analysis: ROC analysis identified a best-performing threshold of 0.8335, while precision-recall analysis identified 0.8334; the study reports an optimal threshold of 0.833.The ROC curve examines the trade-off between true-positive and false-positive rates, whereas the precision-recall analysis focuses on positive-class performance.
- Results: 95.2% accuracy, 90.9% specificity, 95.2% precision, 0.94 ROC area, 0.89 Cohen’s kappa, and 0.96 F1 score were obtained after threshold optimization.The optimized threshold reduced false positives from 4 to 2 and improved accuracy from 92.0% to 95.2%.
4. Explainability of the Proposed Model using LIME
LIME addresses the explainability problem of black-box healthcare models by approximating individual predictions with interpretable local surrogates. For DaTscan images, it uses superpixel perturbations to highlight influential brain regions and support visual interpretation of PD classifications.
- Need for Interpretability: Black-box neural networks in healthcare require explanations so medical workers can assess how diagnostic predictions were reached.The paper frames explainability as important even when a model is accurate, particularly for high-stakes decisions.
- The LIME Model: LIME explains an individual black-box prediction by training an interpretable local surrogate on perturbed samples weighted by their proximity to the case.The surrogate is intended to approximate the original model locally while limiting model complexity.
- The LIME Model: For images, LIME perturbs superpixels by switching connected pixel regions on or off and observing changes in the model prediction.Superpixels can be assigned switching probabilities so the explanation emphasizes the highest-contributing factors.
- Interpretations of DaTscans: The explainer identified the putamen and caudate as high-influence regions, making the diagnostic rationale visually traceable for non-experts.These regions were defined as the study’s region of interest and received the highest explanation weights.
- Interpretations of DaTscans: Healthy-control examples emphasized healthy putamen and caudate regions as influential for non-PD classifications.The corresponding LIME outputs were shown alongside preprocessed SPECT scans.
- Interpretations of DaTscans: PD-classified examples emphasized abnormal or reduced putamen and caudate features as influential regions.The paper presents these visual explanations as corresponding outputs for preprocessed SPECT scans.
5. Conclusion
The study developed a VGG16-based model intended to classify DaTscan images while providing meaningful decision insights. Its reported accuracy supports early diagnosis, but dataset size, class imbalance, labeling reliability, and clinical validation remain important boundaries.
- 95.2% accuracy was achieved using a VGG16 CNN with transfer learning to classify SPECT DaTscan images as Parkinson’s disease or not.
- The model was proposed to make Parkinson’s disease early diagnosis faster, more intuitive, and applicable in real-world scenarios.
- A larger dataset with less class imbalance could support future studies and improve the model’s reliability.
- Model accuracy could be improved through hyperparameter tuning or alternative network architectures, potentially reducing false positives and negatives.
- The study identifies questionable data-label accuracy and the need for clinical validation with novel real-time data.
- The approach is presented as saving healthcare workers’ time and resources while supporting early diagnosis through explanations.