View Proposal
-
Proposer
-
Marta Vallejo
-
Title
-
Advanced Deep Learning and Data Augmentation Strategies for Automated Parkinson's Disease Detection from Digitised Figure-Copying Tasks
-
Goal
-
To investigate how modern deep learning architectures, advanced data augmentation techniques, and explainable artificial intelligence can improve the automated detection of Parkinson's disease from digitised figure-copying tasks while ensuring robust evaluation through subject-wise validation.
-
Description
- Parkinson's disease (PD) is a progressive neurodegenerative disorder whose early diagnosis remains challenging due to subtle motor impairments during the initial stages of the disease. Digitised drawing tasks, such as spiral and pentagon copying, provide an inexpensive and non-invasive source of digital biomarkers capable of capturing motor abnormalities associated with PD.
Recent studies have demonstrated promising results using deep learning for drawing-based PD detection. However, most existing work relies on relatively simple convolutional neural networks and conventional image augmentation techniques. Furthermore, the application of generative models for synthetic data generation remains limited, and there is little understanding of how different augmentation strategies interact with modern deep learning architectures.
The objective of this project is to build upon previous work by systematically investigating advanced artificial intelligence methods for drawing-based PD detection. Depending on the progress of the project, the student may explore several complementary research directions, including:
- comparison of modern deep learning architectures (e.g., ResNet, EfficientNet, Vision Transformers and attention-based models);
- evaluation of transfer learning and self-supervised learning strategies;
- comparison of traditional geometric augmentation with GAN-, DCGAN- and diffusion-based synthetic image generation;
- multimodal fusion of time, pressure and angle representations;
- explainable AI techniques (Grad-CAM, Grad-CAM++, SHAP, Score-CAM) to identify clinically relevant drawing features;
- quantitative assessment of synthetic image quality and its relationship with classification performance;
- investigation of robust validation strategies and the influence of subject-wise versus image-wise dataset partitioning;
- prediction of disease severity or longitudinal progression where suitable datasets become available.
The project aims not only to improve classification performance but also to increase the clinical interpretability, robustness and reproducibility of AI models for Parkinson's disease assessment.
References:
Alissa, M., Lones, M.A., Cosgrove, J., Alty, J.E., Jamieson, S., Smith, S.L. and Vallejo, M., 2022. Parkinson’s disease diagnosis using convolutional neural networks and figure-copying tasks. Neural Computing and Applications, 34(2), pp.1433-1453.
Vallejo, M., Alty, J. and Smith, S.L., 2025. Deep Learning Analysis of Figure Copying Tasks for Parkinson’s Disease Detection with GAN-Based Data Augmentation. medRxiv, pp.2025-04.
- Resources
-
- Existing digitised figure-copying datasets collected from Parkinson's disease patients and healthy controls.
- Previous data representations (time, pressure and angle) developed within the research group.
- Access to previous project code, models and publications developed by the research group.
-
Background
-
Artificial intelligence has become an increasingly important tool for the analysis of digital biomarkers in neurological disorders. Drawing tasks have received considerable attention because they provide objective measurements of motor control while remaining inexpensive and easily deployable within clinical practice.
Previous work from the research group demonstrated that deep learning can successfully discriminate Parkinson's disease patients from healthy controls using digitised figure-copying tasks. That work also showed that traditional geometric augmentation often outperformed more complex GAN-based approaches when using relatively small datasets, highlighting the importance of rigorous methodological evaluation rather than simply adopting increasingly sophisticated algorithms.
Despite recent progress, several important research questions remain unanswered. Little work has systematically compared modern deep learning architectures under different augmentation strategies, and the interaction between transfer learning, attention mechanisms and synthetic data generation remains poorly understood. Similarly, the explainability of drawing-based AI models and the influence of different validation strategies require further investigation before such approaches can be translated into routine clinical practice.
This project will address these challenges by developing robust, interpretable and reproducible AI methodologies for Parkinson's disease detection, with the long-term objective of supporting clinicians through objective, non-invasive decision-support tools.
-
Url
-
-
Difficulty Level
-
Variable
-
Ethical Approval
-
Full
-
Number Of Students
-
2
-
Supervisor
-
Marta Vallejo
-
Keywords
-
parkinson's disease, deep learning, data augmentation, explainable artificial intelligence (xai), computer vision, transfer learning
-
Degrees
-
Bachelor of Science in Computer Science
Master of Engineering in Software Engineering
Master of Science in Artificial Intelligence
Master of Science in Computing (2 Years)
Master of Science in Data Science
Master of Science in Robotics
Master of Science in Software Engineering
Bachelor of Science in Computing Science
Bachelor of Engineering in Robotics
Bachelor of Science in Statistical Data Science
BSc Data Sciences