NUHS-MIT Critical Care Datathon

In the summer of 2018, I participated in an AI Datathon. I signed up for a project that involved analyzing the ISIC 2017 dataset - skin lesion analysis for melanoma detection. Melanoma (skin cancer) presents as irregular and large moles, and this was a dataset full of photos of melanoma mixed with benign moles.

Methods

• Initially, my team and I tried non-deep learning methods like random forests, SVM, but our results were shockingly poor.
• We decided to use deep learning for this computer vision project.
  • We built a basic CNN, but since our images had a high resolution - our GPU kept running out of memory before our model could be trained.
  • We had to think - resizing the image meant that the border of the mole would be lost, and melanoma moles have irregular borders vs non-cancerous moles, which have fairly regular borders.
• We solved this by cutting our image into 4 quadrants.
• We also used different pre-trained models instead of making our CNN from scratch.

Tech Stack

1. Keras
2. Jupyter notebook
3. GPU provided by Google Cloud

Wins

This was my introduction to deep learning - I had no idea about CNNs or computer vision, so I literally had to learn everything overnight. I was helped a lot by my teammates, though - and, thanks to them, we won!

This experience kickstarted my passion for Machine Learning and AI application on healthcare data; so much so that I ended up doing my final year thesis on a similar topic. I also returned to NUH for a part time job, where I created synthetic data for the same hackathon next year.