TL;DR
This paper explores the effectiveness of 1D, 2D, and 3D Convolutional Neural Networks (CNNs) for diagnosing Parkinson's Disease (PD) using drawing tests. They found that increasing the dimensionality of the CNN models, particularly to 3D, leads to significantly improved diagnostic accuracy by better capturing complex kinematic and pressure features of the drawings.
What Problem Does This Paper Address?
Diagnosing Parkinson's Disease (PD) is challenging, and while drawing tests have been used for decades, traditional visual assessment is subjective and lacks quantitative data. Modern approaches leverage digital tablets to capture dynamic handwriting features, but effectively processing this multi-dimensional data for diagnosis remains an active research area. Existing deep learning methods often focus on 2D CNNs applied to image representations of drawings, potentially overlooking valuable temporal and pressure information. This paper aims to bridge this gap by systematically comparing the diagnostic performance of 1D, 2D, and 3D CNNs, hypothesizing that higher-dimensional models can better integrate and leverage the rich dynamic features of handwriting for more accurate PD detection.
Methods
The core of this research lies in transforming raw handwriting data into different dimensional representations suitable for CNN analysis. The authors employed a data enhancement process that encodes various dynamic features (like x-y coordinates, azimuth, altitude, pressure, and velocity) into time series (1D), RGB images (2D), and 3D point clouds.
- Data Enhancement: Raw dynamic features are normalized. For 1D, time series are created directly from coordinates, with velocity replacing timestamps. For 2D, coordinates become pixel positions, and other features (azimuth, altitude, pressure) become RGB color channels, with velocity as line width. For 3D, coordinates and velocity form the (x,y,z) position, and other features become RGB color.
- CNN Architecture: A simplified AlexNet architecture is used across all dimensions, ensuring a fair comparison. The primary difference lies in the convolution kernel and method, which adapt to the 1D, 2D, or 3D input data.
- Data Augmentation: To address the common challenge of limited medical data, techniques like flipping, rotation, illumination, and jitter are applied to augment the datasets.
Datasets
The study utilized two datasets:
- DraWritePD: This dataset was acquired by the authors and contains handwriting data from 25 PD patients and 34 healthy controls (HC). It includes features like x-y coordinates, timestamp, pressure, altitude, and azimuth. The Archimedes spiral drawing test (ASD) was specifically used.
- PaHaW: This publicly available dataset, kindly provided by other researchers, consists of handwriting data from 37 PD patients and 38 HC subjects. It also includes dynamic features captured by a digitizing tablet. The Archimedes spiral drawing test (ASD) was also used from this dataset.
Key Insights and Ideas
- Dimensionality Matters: The central finding is that increasing the dimensionality of the CNN model (from 1D to 2D to 3D) consistently improves diagnostic performance. This suggests that higher-dimensional representations can better capture the complex interplay of kinematic and pressure features indicative of PD.
- Beyond Simple Images: While 2D CNNs applied to image representations of drawings have shown promise, this work highlights that incorporating more dynamic features (like velocity, pressure, and altitude) into higher-dimensional representations (especially 3D point clouds) unlocks superior diagnostic capabilities.
- Feature Encoding is Crucial: The way dynamic features are encoded into the data representation significantly impacts performance. Encoding color information (azimuth, altitude, pressure) and line width (velocity) in 2D and 3D representations proved particularly beneficial.
- Velocity's Nuance: In the 1D case, the velocity feature, while derived from coordinates, surprisingly did not improve predictions over baseline, possibly due to redundancy or how it interacts with the 1D time-series representation. This highlights the importance of how features are integrated into different dimensional spaces.
Key Results
The study reported the following accuracies on the two datasets:
- DraWritePD Dataset:
- 1D CNN: 62.50% accuracy
- 2D CNN: 77.78% accuracy
- 3D CNN: 83.34% accuracy
- PaHaW Dataset:
- 1D CNN: 73.33% accuracy
- 2D CNN: 80.00% accuracy
- 3D CNN: 86.67% accuracy
In both datasets, the 3D convolutional classification method demonstrated the best diagnostic performance. The paper also shows that even with only basic positional information (x,y coordinates), 3D CNNs can achieve competitive results, underscoring their ability to extract meaningful patterns.
Limitations and Future Work
- Feature Set Optimization: While the current feature encoding yields good results, the authors acknowledge that there is still room for improvement. Future research could focus on determining more specific and optimal feature sets to encode.
- Dataset Differences: The paper notes that differences in performance metrics between 2D and 3D CNNs were more pronounced for the DraWritePD dataset compared to PaHaW. This might be influenced by factors like the presence of a reference drawing in DraWritePD, which could make the task easier, and potential differences in drawing direction and age groups of subjects.
- Computational Cost: While not explicitly detailed as a limitation, higher-dimensional CNNs (like 3D) generally incur higher computational costs for training and inference, which is an inherent trade-off for improved performance.
The authors suggest that future work should be directed towards determining specific feature sets to be encoded for even better diagnostic accuracy.
Summary updated