Update from 14/09/22: Published in Applied Soft Computing, February 2022, Volume 116, pages 108261

Update from 03/12/21: To appear in Applied Soft Computing

COVID-Affinity-Model

BibTex citation (preprint):

@article{ter2020one,
  title={One Shot Model For The Prediction of COVID-19 and Lesions 
  Segmentation In Chest CT Scans Through The Affinity Among Lesion Mask Features},
  author={Ter-Sarkisov, Aram},
  journal={medRxiv},
  pages={2020--12},
  year={2020},
  publisher={Cold Spring Harbor Laboratory Press}
}

Bibtex citation (journal publication):

@article{TERSARKISOV2022108261,
title = {One Shot Model For The Prediction of COVID-19 And Lesions Segmentation In Chest CT Scans Through The Affinity Among Lesion Mask Features},
journal = {Applied Soft Computing},
volume = {116},
pages = {108261},
year = {2022},
issn = {1568-4946},
doi = {https://doi.org/10.1016/j.asoc.2021.108261},
author = {Aram Ter-Sarkisov}}

Affinity model

Region of Interest layer - Lesion Mask Features

Affinity matrix for a COVID-19 image. Left column: after 1 epoch, right column: after 100 epochs

Segmentation Results (CNCB-NCOV Segmentation Dataset, (http://ncov-ai.big.ac.cn)

# Affinities	AP@0.5	AP@0.75	mAP@[0.5:0.95:0.05]
32	0.614	0.382	0.395
64	0.603	0.414	0.422
128	0.569	0.350	0.385
256	0.560	0.347	0.386
512	0.548	0.343	0.386

Classification Results (CNCB-NCOV Classification Dataset, (http://ncov-ai.big.ac.cn)

# Affinities	COVID-19	CP	Normal	F1 score
32	89.39%	80.25%	98.96%	90.30%
64	90.68%	83.60%	97.15%	91.00%
128	86.91%	95.65%	95.45%	93.80%
256	91.74%	85.35%	97.26%	91.94%
512	90.27%	84.53%	99.41%	92.34%

Classification Results (iCTCF-CT Classification Dataset, (http://ictcf.biocuckoo.cn)

# Affinities	COVID-19	Normal	F1 score
32	92.11%	80.31%	83.67%
64	86.73%	94.20%	92.00%
128	88.88%	83.85%	85.27%
256	77.41%	93.33%	88.78%
512	90.49%	89.96%	90.11%

Data

CNCB-NCOV data: (ncov-ai.big.ac.cn/download) with COVIDx-CT splits.

iCTCF-CT data: (http://ictcf.biocuckoo.cn/HUST-19.php). Download the nCT(no disease) data. Train and test splits are provided in this repository. I changed the image names to match the convention used in COVIDx-CT: 0 for Negative and 2 for COVID-19.

Testing/Evaluating The Model

To test the model trained on CNCB-NCOV datasets:

python3.5 test_classification_branch.py --ckpt pretrained_models/affinity_model_128.pth --test_data cncb_ncov/test --device cuda --affinity 128

To test the model trained on iCTCF-CT dataset:

python3.5 test_classification_branch.py --ckpt pretrained_models/affinity_model_ictcf_64.pth --test_data ictcf/test --device cuda --affinity 64

This outputs confusion matrix and F1 score above. Links to models are in pretrained_models directory.