# Prediction of total knee replacement using deep learning analysis of knee MRI

## Project Overview

This project implements deep learning models to predict total knee replacement using MRI analysis. Our models achieve up to 0.90 AUC using a combination of MRI and radiograph data, significantly outperforming traditional methods. The implementation supports multiple MRI sequence types (DESS, IW-TSE, T1-TSE) and includes a novel multi-input model architecture.

## Table of Contents

- [Prerequisites](#prerequisites)

- [Installation](#installation)

- [Model Weights](#model-weights)

- [Data Preparation](#data-preparation)

- [Usage](#usage)

  - [DESS Sequence](#dess-sequence)

  - [IW-TSE Sequence](#iw-tse-sequence)

  - [T1-TSE Sequence](#t1-tse-sequence)

  - [Multi-Input Model](#multi-input-model)

- [Results](#results)

- [Citation](#citation)

## Prerequisites

- Python 3.6

- CUDA-compatible GPU (recommended)

- HDF5 dataset files

- Required Python packages (install via `environment.yaml`)

## Installation

1. Clone this repository:

```bash

git clone https://github.com/denizlab/OAI-MRI-TKR.git

```

2. Set up the Python environment:

```bash

conda env create -f environment.yaml

conda activate tkr_mri_env

```

## Model Weights

Pre-trained weights from our 7-fold nested cross-validation are available at:

[https://zenodo.org/records/11237172](https://zenodo.org/records/11237172)

## Data Preparation

### Dataset Requirements

- OAI (Osteoarthritis Initiative) dataset access required

- MOST (Multicenter Osteoarthritis Study) dataset access required

- Contact respective organizations for data access permissions

### Data Organization

- Separate HDF5 files for each MRI sequence type (DESS, IW-TSE, T1-TSE)

- CSV files for dataset splits and cohort information

- Proper file structure for OAI and MOST datasets

## Usage

### DESS Sequence

#### Training

```bash

python3 ./DESS/train.py \

    --file_path /path/to/save/models \

    --csv_path /path/to/split/csvs/DESS \

    --file_folder /path/to/DESS/HDF5/files \

    --val_fold [1-7]

```

#### Evaluation Options

1. Matched OAI Cohort:

```bash

python3 ./DESS/evaluate.py \

    --model_path /path/to/saved/models \

    --csv_path /path/to/split/csvs/DESS \

    --file_folder /path/to/DESS/HDF5/files

```

2. Internal OAI Test Set:

```bash

python3 ./DESS/Eval_OAI_DESS.py \

    --model_path /path/to/saved/models \

    --val_csv_path /path/to/split/csvs/DESS \

    --test_csv_path /path/to/csvs/OAI_SAG_DESS_test.csv \

    --file_folder /path/to/DESS/HDF5/files \

    --vote [True/False]

```

3. MOST Test Set:

```bash

python3 ./DESS/Eval_MOST_DESS.py \

    --model_path /path/to/saved/models \

    --val_csv_path /path/to/split/csvs/DESS \

    --test_csv_path /path/to/csvs/MOST_MRI_test.csv \

    --train_file_folder /path/to/DESS/HDF5/files \

    --file_folder /path/to/MOST/HDF5/files \

    --vote [True/False] \

    --contrast [HR_COR_STIR/SAG_PD_FAT_SAT]

```

### IW-TSE Sequence

#### Training

```bash

python3 ./IW-TSE/train.py \

    --file_path /path/to/save/models \

    --csv_path /path/to/split/csvs/IW-TSE \

    --file_folder /path/to/IW-TSE/HDF5/files \

    --val_fold [1-7]

```

#### Evaluation Options

1. Matched OAI Cohort:

```bash

python3 ./IW-TSE/evaluate.py \

    --model_path /path/to/saved/models \

    --csv_path /path/to/split/csvs/IW-TSE \

    --file_folder /path/to/IW-TSE/HDF5/files

```

2. Internal OAI Test Set:

```bash

python3 ./IW-TSE/Eval_OAI_IWTSE.py \

    --model_path /path/to/saved/models \

    --val_csv_path /path/to/split/csvs/IW-TSE \

    --test_csv_path /path/to/csvs/OAI_SAG_TSE_test.csv \

    --file_folder /path/to/IW-TSE/HDF5/files \

    --vote [True/False]

```

3. MOST Test Set:

```bash

python3 ./IW-TSE/Eval_MOST_IWTSE.py \

    --model_path /path/to/saved/models \

    --val_csv_path /path/to/split/csvs/IW-TSE \

    --test_csv_path /path/to/csvs/MOST_MRI_test.csv \

    --train_file_folder /path/to/IW-TSE/HDF5/files \

    --file_folder /path/to/MOST/HDF5/files \

    --vote [True/False] \

    --contrast [HR_COR_STIR/SAG_PD_FAT_SAT]

```

### T1-TSE Sequence

#### Training

```bash

python3 ./T1-TSE/train.py \

    --file_path /path/to/save/models \

    --csv_path /path/to/split/csvs/T1-TSE \

    --file_folder /path/to/T1-TSE/HDF5/files \

    --val_fold [1-7]

```

#### Evaluation Options

1. Matched OAI Cohort:

```bash

python3 ./T1-TSE/evaluate.py \

    --model_path /path/to/saved/models \

    --csv_path /path/to/split/csvs/T1-TSE \

    --file_folder /path/to/T1-TSE/HDF5/files

```

2. Internal OAI Test Set:

```bash

python3 ./T1-TSE/Eval_OAI_T1TSE.py \

    --model_path /path/to/saved/models \

    --val_csv_path /path/to/split/csvs/T1-TSE \

    --test_csv_path /path/to/csvs/OAI_COR_TSE_test.csv \

    --file_folder /path/to/T1-TSE/HDF5/files \

    --vote [True/False]

```

3. MOST Test Set:

```bash

python3 ./T1-TSE/Eval_MOST_T1TSE.py \

    --model_path /path/to/saved/models \

    --val_csv_path /path/to/split/csvs/T1-TSE \

    --test_csv_path /path/to/csvs/MOST_MRI_test.csv \

    --train_file_folder /path/to/T1-TSE/HDF5/files \

    --file_folder /path/to/MOST/HDF5/files \

    --vote [True/False] \

    --contrast [HR_COR_STIR/SAG_PD_FAT_SAT]

```

### Multi-Input Model

#### Training

```bash

python3 ./MI-DESS_IWTSE/train.py \

    --file_path /path/to/save/models \

    --csv_path /path/to/split/csvs/DESS \

    --file_folder1 /path/to/IW-TSE/HDF5/files \

    --file_folder2 /path/to/DESS/HDF5/files \

    --val_fold [1-7] \

    --IWdataset_csv /path/to/csvs/HDF5_00_cohort_2_prime.csv \

    --DESSdataset_csv /path/to/csvs/HDF5_00_SAG_3D_DESScohort_2_prime.csv

```

#### Evaluation Options

1. Matched OAI Cohort:

```bash

python3 ./MI-DESS_IWTSE/evaluate.py \

    --model_path /path/to/saved/models \

    --csv_path /path/to/split/csvs/DESS \

    --file_folder1 /path/to/IW-TSE/HDF5/files \

    --file_folder2 /path/to/DESS/HDF5/files \

    --IWdataset_csv /path/to/csvs/HDF5_00_cohort_2_prime.csv \

    --DESSdataset_csv /path/to/csvs/HDF5_00_SAG_3D_DESScohort_2_prime.csv

```

2. Internal OAI Test Set:

```bash

python3 ./MI-DESS_IWTSE/Eval_OAI_MI.py \

    --model_path /path/to/saved/models \

    --val_csv_path /path/to/split/csvs/DESS \

    --test_csv_path1 /path/to/csvs/OAI_SAG_TSE_test.csv \

    --test_csv_path2 /path/to/csvs/OAI_SAG_DESS_test.csv \

    --file_folder1 /path/to/IW-TSE/HDF5/files \

    --file_folder2 /path/to/DESS/HDF5/files \

    --vote [True/False] \

    --IWdataset_csv /path/to/csvs/HDF5_00_cohort_2_prime.csv \

    --DESSdataset_csv /path/to/csvs/HDF5_00_SAG_3D_DESScohort_2_prime.csv

```

## Results

### Model Performance

**Table 1: Receiver operator characteristic analysis with areas under the curve (AUC) and area under the precision-recall curve (AUPRC) evaluating the diagnostic performance of the models to predict total knee replacement (TKR) using sevenfold nested cross-validation on the training a validation group in the OAI database.**

| Model | AUC (95% CI) | p value | AUPRC (95% CI) | Sensitivity (%) (95% CI) | Specificity (%) (95% CI) |

|-------|--------------|----------|----------------|-------------------------|-------------------------|

| **MLP model** |

| Traditional | 0.77 (0.74, 0.81) | Reference | 0.76 (0.71, 0.81) | 73 (68, 77) | 73 (68, 78) |

| **CNN models** |

| DESS | 0.88 (0.86, 0.91) | <0.001 | 0.87 (0.83, 0.91) | 82 (78, 86) | 81 (77, 85) |

| FS-IW-TSE | 0.86 (0.84, 0.89) | <0.001 | 0.87 (0.84, 0.90) | 77 (73,82) | 84 (80, 87) |

| Multi-input MRI | 0.85 (0.82, 0.88) | <0.001 | 0.85 (0.81, 0.89) | 79 (75, 83) | 79 (74, 83) |

| IW-TSE | 0.87 (0.84, 0.90) | <0.001 | 0.87 (0.84, 0.90) | 82 (78, 86) | 78 (73, 82) |

| Radiograph | 0.87 (0.84, 0.89) | <0.001 | 0.87 (0.84, 0.90) | 81 (76, 85) | 80 (76, 84) |

| **Ensemble models** |

| MRI | 0.89 (0.87, 0.91) | <0.001 | 0.89 (0.87, 0.91) | 79 (75, 83) | 86 (82, 89) |

| MRI and radiograph | 0.90 (0.87, 0.92) | <0.001 | 0.90 (0.87, 0.93) | 80 (76, 84) | 85 (81, 88) |

Key findings from our results:

- All CNN and ensemble models significantly outperformed the traditional MLP model (p < 0.001)

- The MRI and radiograph ensemble model achieved the highest performance with AUC of 0.90 (95% CI: 0.87-0.92)

- DESS sequence alone showed strong performance with AUC of 0.88 (95% CI: 0.86-0.91)

## Citation

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

[https://www.nature.com/articles/s41598-023-33934-1](https://www.nature.com/articles/s41598-023-33934-1)

```bibtex

@article{rajamohan2023prediction,

   title={Prediction of total knee replacement using deep learning analysis of knee MRI},

   author={Rajamohan, Haresh Rengaraj and Wang, Tianyu and Leung, Kevin and Chang, Gregory and Cho, Kyunghyun and Kijowski, Richard and Deniz, Cem M},

   journal={Scientific reports},

   volume={13},

   number={1},

   pages={6922},

   year={2023},

   publisher={Nature Publishing Group UK London}

}

```