I appreciate the opportunity to be a part of Dr Joseph Jacob’s team working on “Automatic segmentation and analysis of cystic fibrosis lung CT scans”. This unique experience was made possible through the Summer Studentship award granted by BALR. I worked on this project for 6 weeks at the UCL Centre for Medical Image Computing (CMIC), straight after completing my third year of medical school and my iBSc in Computing, Maths and Medicine.
The basic aim of the project is to develop a deep learning algorithm capable of processing CT scans from CF patients and generating precise 3D models of the airways along with quantifiable metrics like airway dilatation and plugging volume. To start, I familiarised myself with the Linux command-line interface. This operating system allowed to me to readily handle thousands of CT image files very efficiently. I was tasked to work through a large hospital dataset of CF lung CT scans.
I meticulously examined and categorized approximately 600 CF lung CT scans, with diverse disease morphologies. Through hands-on experience and guidance from my supervisor, I honed my ability to identify various disease patterns, including atelectasis, honeycomb cysts etc. This significantly enhanced my pattern recognition skills, which will be invaluable for any future medical segmentation projects I may embark upon.
After categorising all the scans, I utilised python to extract relevant scans based on our inclusion criteria. I ran our initial nnU-net model these scans to produce output segmentations. The training of the initial deep learning model was very resource-intensive and required 3 days on a high spec GPU workstation. Volumetric calculations were made on the airway dilatation, airway mucus plugging and total lung volume. Accuracy was assessed using various metrics that were calculated, such as DICE Score and percentage pixel loss after post-processing. I then sought to improve the performance of this deep learning model by further manually refining some of the output segmentations and feeding them back into the original. I used the Student’s t-T
est to determine whether there was any significant difference in the change in accuracy metrics. In the future, we’d like to correlate the airway dilatation/plugging volumes with current clinical outcome measures to observe whether this would make a suitable or improved clinical endpoint.
Throughout this studentship, I leveraged and expanded upon my programming and statistical analysis skills acquired during my BSc, honing their application within a research-oriented context. I also learnt the value of resilience in research, as I confronted and systematically resolved numerous errors encountered during the project. Towards the end of my project, I learnt how to use python to create bar charts and boxplots from my data offering an easily comprehensible means of presenting my results to my supervisor.
Furthermore, I was very grateful to be given the chance to carry out my work in the office, where I could interact with fellow colleagues working on similar projects and share ideas. Hearing their experiences proved immensely beneficial, offering me unique insights into their research pursuits, which helped clarify my own aspirations for a future career in clinical research.
All in all, this was an invaluable experience for me, and I would like to thank the BALR for giving me the chance to contribute to this project and for allowing me to contemplate a potential career in clinical research. My supervisor Dr Gabrielle Baxter also deserves a special thanks for her excellent support and guidance. The knowledge I acquired working in this lab surpasses what can typically be obtained from remote work or lectures.