An Ordinal Regression Framework for a Deep Learning Based Severity Assessment for Chest Radiographs

In this repository we share our code from the identically named paper. You can find it using the following link: https://arxiv.org/abs/2402.05685 . Unfortunately, we are not allowed to share the clinical dataset we used.

Requirements

module
torch
torchvision
torchmetrics
pytorch_lightning
PIL
wandb

---

Run the code (example)

    python3 argParser.py \
    --batchSize 32 \
    --learningRate 0.005 \
    --numEpochs 30 \
    --netModel resnet50 \
    --fiveFold 0 \
    --targetFunc gauss \
    --classFunc "l1dist;dotProd" \
    /

Argument	Type	Description
--batchSize	int	Required. Batch size
--learningRate	float	Required. Learning rate
--numEpochs	int	Required. Number of epochs
--netModel	str	Required. Used model (e.g. resnet50, deit)
--fiveFold	int	Required. Fold of five fold cross validation
--targetFunc	str	Required. See below
--classFunc	str	Required. See below

Target Function

The target function can be specified with the --targetFunc argument.

--targetFunc	Description
gauss	Gaussian encoding
oneHot	One-Hot encoding
continuous	Continuous encoding
progBar	Progress-Bar encoding
softProg	Soft-Progress-Bar encoding
binNum	Binary-Number encoding

The exact mappings are defined in targetMap.py.

Classification Function

The target function can be specified with the --classFunc argument. Since the training does not depend on the classification function, several can be evaluated on the same training. To do this, separate the different arguments for the classification function with a ;.

--classFunc	Description
argmax	Argmax function
sum	Sum of vector
l1dist	L1 distance
dotProd	Normalized dot product

The implementation of this can be found in classMap.py

How to cite our work

If you find the paper or code useful for your academic work, please consider citing our work.:

[TODO ADD BIBTEX]