Using radiomics features to predict molecular subtype of breast cancer

BDSI Imaging Subgroup

Project Overview

This project focuses on using radiomics features extracted from medical imaging to predict molecular subtypes of breast cancer. We explore a variety of linear models, custering methods, and machine learning models as well as subsets of features to determine which combination of model and feature may yield the best performance.

Methodology

Our research employs a comprehensive approach to analyze three critical cancer biomarkers:

  • ER (Estrogen Receptor) - Key hormone receptor in breast cancer
  • HER2 (Human Epidermal Growth Factor Receptor 2) - Important therapeutic target
  • PR (Progesterone Receptor) - Hormone receptor affecting treatment decisions

Model Zoo

Comprehensive evaluation of multiple machine learning algorithms including ensemble methods, neural networks, and traditional classifiers.

Performance Metrics

Detailed AUC analysis across five-fold cross-validation and feature importance evaluation across all biomarkers, as compared to previously published baselines.

Visualizations

Heatmaps, box plots, and bar charts showcasing model performance and feature relationships.

Explore Results

Research Results

Comprehensive analysis results for cancer biomarker prediction models, as compared to previously published baselines in A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features..

Estrogen Receptor (ER) Analysis

AUC Performance by Model
ER AUC Bar Plot
AUC Distribution
ER AUC Box Plot
Performance Heatmap
ER AUC Heatmap

HER2 Analysis

AUC Performance by Model
HER2 AUC Bar Plot
AUC Distribution
HER2 AUC Box Plot
Performance Heatmap
HER2 AUC Heatmap

Progesterone Receptor (PR) Analysis

AUC Performance by Model
PR AUC Bar Plot
AUC Distribution
PR AUC Box Plot
Performance Heatmap
PR AUC Heatmap