nicemorning / COMP683-S24

Introduction

This repository hosts the COMP683 final project: Tumor Type Prediction using Convolutional Neural Networks (CNNs). The classification results for 33 tumor types are summarized in the following table.

CNN Architecture	# Parameters	Test Accuracy (%)
ResNet18	11,193,441	95.30
ResNet50	23,575,649	94.85
MobileNetV2	2,266,145	94.30
SqueezeNet	739,425	91.14

Repository Structure

• src/: Contains all source files.
  • common_functions.py: Includes common functions for CNN training and testing.
  • image_transformer.py: Transforms 1-D features to 2-D images using feature similarity, adapted from DeepInsight.
  • process_tcga_dataset.py: Reads TCGA RNA-seq data for 33 tumor types and converts the sample-feature matrix into 2-D images.
  • test_cnns.py: Trains and evaluates prediction accuracy across 33 tumor types using various CNN architectures on converted 2-D images.
  • tcga_files.py: Lists predefined filenames and tumor names for the 33 tumor types, sourced from the UCSC Xena Portal.

How to Reproduce the Results

1. Download TCGA RNA-Seq Data

• Visit the UCSC Xena Portal.
• Select the Harmonized dataset by clicking on the GDC hub checkbox.
• Download the required files listed in ./src/TCGA/tcga_file_list.txt into the ./src/TCGA folder

2. Convert Gene Expression Data to 2-D Images

Navigate to the src directory, run the following Python script to convert the TCGA gene expression data into 2-D images:

python process_tcga_dataset.py

This script will generate a file named feat2img.bin, which includes all the 2-D images along with their corresponding labels.

3. Train and Evaluate CNN Models

Continue in the src directory, execute following Python script to train and evaluate the prediction accuracy for 33 tumor types using various CNN architectures, including SqueezeNet, MobileNet, ResNet18, and ResNet50:

python test_cnns.py