DL4H-SP24 Local Explanations For Cervical Cancer

This project is a reproduction of the work done in Local Explanations for Cervical Cancer, developed as part of the UIUC Deep Learning for Healthcare Course by Team 11: Logan Borders, Ankita Singh, and Sam Kuhbander. Video: https://drive.google.com/file/d/1VbnpcCaYxyATlkzQZPA5bLop64rc6XSg/view?usp=sharing

Prerequisites

The data access information is available in the "Data" section, and all dependencies / necessary downloads are show in code.

Usage

The checkpoints are labeled as such in the headers to make it easier for you to run, but everything has been commented out that is not necessary. For pandas pip install conflicts you are able to hit cancel on warnings

Data

Accessing the Data

• Google Drive Folder: Access the Data Here
  • Please contact Logan, Ankita, or Sam if you encounter any access issues.

Computational Requirements

• Hardware Used: The dataset is small enough for computations on an 8 GB 2133 MHz LPDDR3 2.3 GHz Dual-Core Intel Core i5.
• Run Time: Average training time is less than 2 minutes.
• Epochs: For logistic regression, default epochs are 100 and for MLP it is 200, unless convergence is reached earlier.
• The training for the Local explainability models took hours however and we had to sample to get them done.

Notebooks

• Draft Submission: View the Draft Notebook

  • Instructions: Ensure you "Run all" in the notebook. The checkpoints will automatically handle data processing and model training.

• Full Code: View the Full Code Notebook

Results

The study replicated findings on the Random Forest Classifier's accuracy in predicting cancer and the LIME method's effectiveness in explaining predictions. Results mostly matched the original paper, showing Random Forest's reliability and LIME's superior explanation accuracy.

Contact

For any issues with access or questions regarding the process, please contact:

• Logan Borders
• Ankita Singh
• Sam Kuhbander