EN 250 Final Project

Detect malaria using NIH parasitized cell slide image dataset and the Google EfficientNet architecture.

Directories

Directory	Description
results	Results of training the model on an AWS EC2 "p3.2xlarge" instance.
results/by_batch_size	Graphs of accuracy and loss for each minibatch size tested, as well as two graphs that contain all those results combined.
results/final_model	The trained, full model.

The model included in the results is not the divergent model used in the paper and presentation, instead it is a model which properly converged. The solution to the issue of divergence in the original model was the use of image preprocessing and a lack of available data. Data augmentation and removal of image preprocessing resolved the convergence issues to produce a model which achieves 95.8% accuracy with a categorical cross-entropy loss of 0.128.

Setup

1. Install Anaconda (https://www.anaconda.com/products/individual)
2. conda create --name tfgpu tensorflow-gpu
3. conda activate tfgpu
4. pip install -U git+https://github.com/qubvel/efficientnet
5. pip install -U scikit-learn
6. Navigate to be inside the repository directory
7. wget ftp://lhcftp.nlm.nih.gov/Open-Access-Datasets/Malaria/cell_images.zip
8. unzip cell_images.zip
9. mv cell_images/ data/
10. python malaria.py