Assessment of knee pain from MR imaging using a convolutional Siamese network

This work is published in European Radiology (https://doi.org/10.1007/s00330-020-06658-3).

Introduction

Prerequisites

The tool was developed based on the following dependencies:

1. PyTorch (1.1 or greater).
2. NumPy (1.16 or greater).
3. Scipy (1.30 or greater)
4. OpenCV (3.4.2 or greater)
5. scikit-learn (0.21.2 or greater)

Please note that the dependencies require Python 3.6 or greater. We recommend installation and maintenance of all packages using conda. For installation of GPU accelerated PyTorch, additional effort may be required. Please check the official websites of PyTorch and CUDA for detailed instructions.

Data files

Original imaging file in .npy stored in:

data/raw/NAME_OF_SEQUENCE/*.npy

Registered imaging file in.npy stored in:

data/registered/NAME_OF_SEQUENCE/*.npy

Name of sequences contain

SAG_IW_TSE_LEFT & SAG_IW_TSE_RIGHT

Examples

To start a new training job with learning rate of 1e-4, batch size of 64, and learning rate decay of 0.9

$ python main.py --lr 0.0001 --bs 64 --gamma 0.9

To perform linear registration

$ python main.py --r

To continue model training from a saved checkpoint:

$ python main.py --c

Arguments

1 . --rp:

Run using multiple GPUs in parallel.

2 . --r:

Run along with linear registration.

3 . --c:

Run using the saved checkpoints.

4 . --lr: float

Learning rate, default is 1e-4

5 . --gamma: float

Learning rate decay, default is 1.0

6 . --epochs: int

Number of epochs, default is 500

7 . --bs: int

Batch size, default is 64

Authors

• Gary Han Chang, [email protected] - Kolachalama laboratory, Boston University School of Medicine