Joshua Chan

I was involved in the workstream of medical image segmentation, focusing on producing semantic segmentation masks for Chest X-ray data. The predicted masks data would then be imported into a labelling software. The team’s ultimate goal was to facilitate radiologists to produce manual-labelled segmentation masks more efficiently, which would become part of the data input for future model comparisons.

I gained a very steep learning curve from this project. In my first week, I recalled that I was not even familiar with some of the basic libraries in Python (i.e., cv2, matplotlib, numpy, pandas), and I had to look some reference code from GitHub and Kaggle online. Into the first three weeks of internship, I even had to rush through a deep learning course on Coursera in order to pick up the basic concepts of CNN and U-Net - the neural network architecture for image segmentation tasks. Luckily, I managed to learn it quickly, and during the fourth to fifth  week, I was able to build an initial product prototype and complete the product in the last week of internship.

While my project was more product-based, I also did some literature reviews on domain adaptation and data distribution shift problems. This was inspired by some of the observations on the predicted images. However, the literature I read was pretty theoretical, and I wasn’t able to leverage the literature findings to improve my model. On the other hand, the conventional methods of changing the model hyper-parameters, such as the learning rate and the loss function, improved the model drastically.

The entire project team included my supervisor, Dr. Xu Xinxing, Scientist, Yi Pin, a full-time researcher, and another intern - Duong Ngoc Yen. Although my focus was on image segmentation tasks, I would also discuss with other teammates and handle some ad hoc tasks where help was needed.

The work-from-home mode throughout this internship gave me a lot of flexibility in allocating my time of the week. Dr. Xu, Yi Pin, Yen and myself would gather to discuss our work progress weekly. During this meeting, we would discuss different projects that different team members simultaneously undertook, which would give me a broader picture of the entire team’s workflow.

The group meetings usually took place in the middle of the week, i.e., Tuesday or Wednesday. Using this meeting as a mini-deadline to achieve specific weekly progress, I would usually allocate most of the effort at the beginning of the week (i.e., Monday to Wednesday) to finalise my progress and generate results to report in the meeting. Thus, it would be the busiest moment of the week. Then I would spend the rest of the week doing some literature review or taking up some ad hoc tasks as requested by the team.

Q: How has this internship benefits you?

As I plan to do a Ph.D. in the future, the first-time research experience that I gained from this internship would benefit my future planning. Specifically, it gave me confidence that I could pick up a steep learning curve in a short amount of time and could self-taught and self-learn complex concepts through reading research papers. In addition, some of the ad hoc tasks that I worked on with my team members also trained me how to work dynamically as a team. 

Besides, an NTU professor also invited me to undertake a URECA (undergraduate research) project with him for the next academic year, after displaying the research skills I gained from IHPC. I am now very determined that computational research will be my career choice in the near future.