Subchondral bone length in knee osteoarthritis: A deep learning driven imaging measure and its association with radiographic and clinical outcomes

This work is published in Arthritis & Rheumatology (https://doi.org/10.1002/art.41808).

Prerequisites

The tool was developed based on the following dependencies:

1. PyTorch (1.1 or later).
2. PyTorch Lightning (1.3 or later).
3. NumPy (1.16 or later).
4. Scipy (1.30 or later)
5. scikit-learn (0.21.2 or later)
6. Tensorboard

Project Structures

├── checkpoints              # Saved checkpoints of the U-Net models
├── data                     
    ├── dess_annotated       # Annotated DESS MRI images for training of the neural network
        ├── img                MRI images in .jpg format
            ├── 1_10.jpg
            ├── ...
        ├── masks
            ├── fc             Annotated masks of femur cartilage (fc), medial tibia cartilage (mtc), lateral tibia cartilage (ltc), etc...
                ├── 1_10.jpg
                ├── ...
            ├── mtc
            ├── ltc
            ├── ...
    ├── testing              # DESS MRI for predictions
├── dess_utils               # Tools to create segmentations and to perform statistical analysis
    ├── SBL_statistics.py    # Statistical analysis of SBL data
├── engine                   # Engine of Pytorch Lightning
├── loaders                  # Loaders for knee DESS MRI images and annotations
├── logs                     # Training logs of Pytorch
├── models                   # Definition of Pytorch models
├── ln_segmentation.py       # main script to perform model traning of bone and cartilage segmentation
├── prediction.py            # calculate SBL based on predicted bone and cartilages masks
├── postprocess.py           # cleaning segmentation results, create SBL based on predicted bone and cartilages masks
└── README.md

Segmentation of MRI-Based Knee Shape

Data files

Pairs of images and masks of segmentation should be put in:

data/CASE_NAME/train_imgs/*.png
data/CASE_NAME/train_masks/*.png

for training dataset and

data/CASE_NAME/eval_imgs/*.png
data/CASE_NAME/eval_masks/*.png

for validation dataset. The images for testing the segmentation results should be put in:

data/CASE_NAME/test_imgs/*.png

And the segmentation results will be created in:

data/CASE_NAME/test_masks/*.png

Neural Network Training of Automatic Segmentation

python ln_segmentation.py

Sample Scripts for prediction of bone and cartilage segmentation

python prediction.py

Sample Scripts for Postprocessing

python postprocess.py

Stastistics

Demographics

Subchondral Bone Length Measurement

Odds Ratio

Box plots

SBL Difference Plot

Others

Run script

cd ~/boneshape/sbl_OR/

python3 sbl_boxplot.py

Prerequisites

SBL data in ~/boneshape/df_extracted/SBL_0904.csv. Each row corresponds to a subject knee. merge1.csv in ~/boneshape/df_extracted/merge1.csv contains the clinical data. Each row corresponds to a patient, while each column corresponds to demographic/clinical info. Index order must be maintained in merge1.csv and SBL_0904.csv.

Dependencies include:

1. pandas version 1.1.5
2. numpy version 1.19.5
3. scipy version 1.5.4
4. matplotlib version 3.3.4
5. seaborn version 0.11.1
6. sklearn version 0.24.1
7. colorsys

Results

sbl_boxplot.py script will produce the odds ratio data, box plots, and, SBL Difference Plot. Generated figures are located in ~/boneshape/figures.