Preparing Spiral Drawings for Parkinson's Screening
Why preprocessing matters
A machine-learning model can only learn from the signal preserved in its input. For spiral drawings, inconsistent image sizes, background noise, and differences in contrast can make comparisons harder.
My final-year project explores Parkinson's disease screening from handwriting and spiral-drawing images. I am treating preprocessing as a separate, testable stage rather than mixing it into model training.
The current pipeline
The preprocessing work uses Python and OpenCV to standardize input images. The pipeline includes:
- Converting inputs to a consistent grayscale representation.
- Applying thresholding to separate the drawing from the background.
- Using denoising where it helps preserve the stroke.
- Resizing outputs to a consistent image size.
- Checking that the resulting images remain valid for downstream feature extraction.
Otsu thresholding is a useful starting point, while adaptive thresholding can help when lighting or background intensity varies. Neither should be assumed to work best for every sample without visual inspection.
What comes next
The next stage is feature extraction. Candidate features include shape descriptors, texture features, frequency-domain features, and measurements specific to spiral geometry.
I will compare these features using a consistent evaluation setup rather than treating a single accuracy score as proof of clinical usefulness.
A note on limitations
This is an academic screening experiment, not a diagnostic tool. Dataset quality, participant-level data leakage, class balance, and validation design all affect how trustworthy the results are. Any reported performance needs to be measured on held-out data and described with those limitations.
Takeaway
Keeping preprocessing reproducible makes it easier to inspect failures, compare feature extraction methods, and understand what information the model is actually receiving.