This is a small computer vision study comparing 2D vs 3D CNN architectures in classifying lung disease (TB) from chest CT scans. More specifically, I compare three CNN models of increasing complexity (baseline, + batch normalisation, ++ more layers) in two parallel experiments, 2D and 3D equivalents. In the end, I compared 6 in-house designed CNN's, and out of curiosity also included a RESNET50 transfer learning model.
- Python Scripts & Notebook: Python scripts & notebooks!
- 2D Models: python scripts showing 3x 2D CNN designs.
- 3D Models: python scripts of 3x 3D CNN models - these are the 3D equivalents of the above 2D models.
- MosMedCT_EDA.ipynb: Exploratory analysis and pre-processing of this CT dataset.
- Run_Models.ipynb: This is where I use Sagemaker notebook to deploy all these models in an EC2 instance in AWS.
- README.md: Hello, you're reading me.
- Report.pdf: An academic-style report explaining the motivation of the study and reporting the results. Spoiler alert, 3D wins, but the final results may still surprise you, so I encourage you to have a look at the pdf report.