Breast-Cancer-MRI-Classification

This repository contains our TDT4265 mini-project on breast cancer classification from DCE-MRI in the ODELIA 2025 challenge.

Overview

Breast cancer classification from MRI is challenging because the data are:

• high-dimensional 3D volumes,
• strongly imbalanced across classes,
• and affected by domain shift across hospitals.

Our project investigates both 3D volumetric models and 2D MIP-based models, and uses logit ensembling to improve generalization on unseen centers.

Data Availability

The dataset used in this project belongs to the ODELIA 2025 challenge and is not included in this repository due to access and privacy restrictions.

Method

Our final submission combines:

• a 3D ResNet18 trained on volumetric MRI inputs,
• a 2D ResNet50 trained on Maximum Intensity Projection (MIP) images,
• test-time augmentation (TTA),
• and logit-space ensembling.

Final models

• 3D ResNet18
  • input modalities: Pre, Post_1, Post_2
  • isotropic resampling
  • background masking
  • light augmentation
• 2D MIP ResNet50
  • MIP channels: Post_1, Sub_1, Sub_2, Post_2
  • weighted and unweighted variants
  • ensemble at inference time

Main Finding

A key result of this project was that averaging logits performed better than averaging probabilities.
This improved leaderboard performance without changing the underlying models.

Results

• Final leaderboard score: ~0.57
• Final rank: 9 / 28 teams
• Best single 3D model: ~0.54
• Logit ensemble improvement: about +0.02

Repository Structure

src/
├── analysis/
│   └── eda.py
├── baselines/
│   ├── densenet121_192.py
│   ├── densenet169_384.py
│   ├── efficientnet_192.py
│   ├── train_densenet121_384.py
│   ├── train_densenet169_384.py
│   └── train_efficientnet_384.py
├── final_models/
│   ├── resnet18_3d.py
│   └── mip_resnet50_ensemble.py
└── inference/
    └── ensemble_inference.py