Being nearly a year past after coming back from Taiwan, I write here the gist of the project to share and record my experience.

This post is about the project that I did with one friend from Germany and another friend from the Philipines doing his doctorate in NTU for IE graduate course in NTU.

Research Objective

After the first outbreak of COVID-19, countries had trouble due to the lack of the diagnosis kit and the slow speed of the process. This project tries to complement this problem by analyzing the CT scans of the patients.

Data Source

A precedent research and a Kaggle project offered available dataset.

Data Preprocessing

Dimension Reduction

To find the optimal reduction parameter, we selected potential candidates and drew variables to work on.

Using Statitical Models

On the new transformed and non-transformed dataset, we used various statistical machine learning methods to find out if we can use CT scans to detect COVID-19.

Demension reduction techniques do not accurately discern the areas with COVID-19. This is because lesions occur on random spots, meaning that we might be able to get better results by using NN based models.

Using a NN Based Model

We used transfer learning on ResNet50 pretrained with Imagenet. Data Augmentation was also done.

Implementation of Transfer Learning on ResNet50

Using Tensorboard for Training

Visualization of the weights of a pretrained ResNet

Small amount of change happens during fine tuning.

Training the top layer for 5 times is better than traning for 8 times.


Implementation of the Training Process

DL vs. ML

DL overall better in terms of accuracy.

No reason to sacrifice accuracy in this case. DL is better.

Both the ML and the DL does not provide interpretability. DL is a overall blackbox model, and the the dimension reduction technique used in ML cannot specify the areas of lesion.

Despite the room for improvement, the potential to be used as a inexpensive measure of COVID-19 detection is present.