MRI Sequential Active Sampling

This repository contains the implementation of a sequential active sampling approach for MRI diagnosis. The project focuses on optimizing MRI acquisition by intelligently selecting the most informative samples to achieve accurate classification with reduced scan time.

Overview

MRI scans traditionally require long acquisition times to gather sufficient data for accurate diagnosis. This project implements machine learning techniques to:

1. Select optimal k-space measurements (active sampling)
2. Minimize the number of MRI acquisitions needed for accurate diagnosis
3. Evaluate different sampling strategies for specific knee conditions and severity degree.

Project Structure

The repository is organized into two main components:

1. Classifier: Models for classifying knee conditions from MRI data
2. Weighted_Sampler: Implementation of weighted sampling policies for active learning

Requirements

click==8.1.7
h5py==3.11.0
lightning==2.3.3
natsort==8.4.0
numpy==2.1.0
pandas==2.2.2
pytorch_lightning==2.3.3
PyYAML==6.0.2
scikit_learn==1.5.1
torch==2.3.1
torchmetrics==1.4.0.post0
torchvision==0.18.1
wandb==0.17.7

Installation

1. Clone this repository:

   git clone https://github.com/your-username/MRI_Sequential_Active_Sampling.git
   cd MRI_Sequential_Active_Sampling

2. Install the required dependencies:

   pip install -r requirements.txt

Usage

Classifier Experiments

Run various classifier experiments with different configurations and random seeds:

bash run_classifier_experiments.sh

This script will execute training for the following configurations:

• ACL Sprain detection (binary classification)
• Cartilage Thickness Loss detection (binary classification)
• ACL Sprain degree classification (binary classification)
• Cartilage Thickness Loss degree classification (binary classification)

Each experiment is run with 5 different random seeds (42, 123, 456, 789, 1024) to ensure robust evaluation.

Weighted Sampling Experiments

Train weighted sampling policies for active learning:

bash run_weighted_experiments.sh

This script trains weighted policies for:

• ACL Sprain detection and degree quantification
• Cartilage Thickness Loss detection and degree quantification

Configuration

Configuration files are available in the Classifier/config/ and Weighted_Sampler/config/ directories:

• knee_acl_config.yaml: Configuration for ACL Sprain detection
• knee_cart_config.yaml: Configuration for Cartilage Thickness Loss detection
• knee_acl_degree_config.yaml: Configuration for ACL Sprain degree classification
• knee_cart_degree_config.yaml: Configuration for Cartilage Thickness Loss degree classification

Experimental Setup

The experiments use the following parameters:

• Initial accelerations: 5% (starting with fewer measurements)
• Final accelerations: 30% (ending with more measurements)
• Multiple random seeds for statistical validity

Tracking

This project uses Weights & Biases (wandb) for experiment tracking and visualization.

Citation

If you use this code in your research, please cite:

@misc{du2025activesamplingmribasedsequential,
      title={Active Sampling for MRI-based Sequential Decision Making}, 
      author={Yuning Du and Jingshuai Liu and Rohan Dharmakumar and Sotirios A. Tsaftaris},
      year={2025},
      eprint={2505.04586},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.04586}, 
}