PRISM: Plain MRI Recognition and Interpretable System for hepatic Malignancy

PRISM is a state-of-the-art, multi-stage deep learning framework designed for automated, gadolinium-free diagnosis of Focal Liver Lesions (FLLs) using only non-contrast MRI (NC-MRI).

Developed and validated on a massive multicenter cohort of 12,823 patients from 9 institutions, PRISM addresses the limitations of dynamic contrast-enhanced MRI (DCE-MRI) by providing a fast, non-invasive, and expert-level diagnostic tool.

✨ Features

• State-of-the-Art Model: Utilizes the powerful Uniformer-B architecture, designed for temporal and volumetric data.
• Cross-Validation: Implements a robust 5-fold cross-validation strategy for training and evaluation to ensure model generalization.
• End-to-End Workflow: Provides clear, executable scripts for the entire pipeline: data preparation, training, inference, and model ensembling.
• Reproducibility: Includes clear instructions to set up the environment and run the experiments.

🚀 Getting Started

Follow these instructions to set up and run the project on your local machine.

Prerequisites

• Python 3.8+
• PyTorch
• pip for package management

Installation

1. Clone the repository:

   git clone https://github.com/puyln/PRISM.git
   cd PRISM

2. Install the required dependencies:

   pip install -r requirements.txt

⚙️ Usage Workflow

The project is structured into a simple, three-step process.

1. Data Preparation

• Input Data: This model is designed to work with lesion-centered 3D ROIs. Place your dataset in the appropriate directory structure as referenced by the scripts, likely under ./data/classification_dataset.
• Data Splits: We provide pre-defined data splits for 5-fold cross-validation. You can find the label files in ./data/classification_dataset/labels. You can also generate your own splits as needed.

2. Model Training

To train the models, you first need the base pre-trained weights for Uniformer-B.

• Download Pre-trained Model:

  1. The model is initialized with weights from the official Uniformer-B model, pre-trained on the Kinetics-400 dataset. You can find the official repository here: Sense-X/UniFormer Official GitHub
  2. You will need to download the specified model version (Kinetics-400, #Frame:8x1x4, Sampling Stride:8).
  3. Since the downstream task has a different classification head, you must remove any layers with mismatching shapes from the weight file (e.g., the final classification layer) to create a "pruned" version.
  4. Place the processed weight file in the ./pretrained_weights/ directory.

• Run the Training Script: Once the pre-trained model is in place, you can start training the two sets of 5-fold cross-validation models (10 models in total) using the following command:

  sh ./main/train.sh

  The training progress and resulting model checkpoints will be saved in their respective directories.

3. Prediction and Ensembling

• Run Inference: Execute the prediction script to generate scores from all 10 trained models.

  sh ./main/result.sh

  This will generate prediction files in the subfolders of ./results/.

• Ensemble the Results: Finally, merge the scores from all prediction files to produce a single, more robust prediction.

  sh ./main/ensembling.sh

  The final ensembled prediction file will be created at ./results/merged_score.json.