Projects
Honours thesis
Automated analysis of cochlear implant CT scans. Awarded the highest mark in the course and the University Medal.
Problem
Cochlear implant surgery is largely a blind procedure. Surgeons insert a flexible electrode into the cochlea with limited visibility and tactile feedback. CT imaging is the gold standard for checking electrode position after insertion, but correctly orienting the scan and identifying the relevant features was a tedious manual process that could take a skilled operator 10 to 20 minutes per case.
Existing automated CT-analysis tools were limited in extensibility and clinical generality. Many were MATLAB-based, expensive to run, and tested only on narrow sets of high-quality images. The research question was how to build an automated, extensible tool for clinical CT data that could obtain the cochlear view and locate each electrode pad and the modiolar axis.
What I did
I built a Python library and API for the automated analysis of cochlear implant CT scans. The pipeline combined numerical and machine-learning methods to segment the implant, rotate the scan into the cochlear view, identify the electrode path, interpolate the 22 electrode positions, and estimate the modiolar axis.
I compared four segmentation methods, including Otsu thresholding, k-means and modal segmentation, and an annealing-inspired optimisation approach. I used PCA and affine transformations to generate the cochlear view, created a pixel adjacency graph to trace the electrode spiral, and experimented with Euclidean distance residual minimisation to fit Cohen and modified Cohen spirals.
Outcome
The tool reduced the cochlear-view workflow from a manual process of roughly 10 to 20 minutes to less than 10 seconds per image. Modal and k-means segmentation performed best for clipped scans, while Otsu performed best for non-clipped scans, and the cochlear-view algorithm was robust across varied clinical CT datasets.
Electrode localisation worked well across many images. Movement, starburst artefacts, and highly variable insertions showed the limits of simple curve-fitting. The project was awarded the highest mark in the course and the University Medal.