Update 28/12/20

I evaluated SSM on the test split of iCTCF-CT data (http://ictcf.biocuckoo.cn), 12976 CT slices, here's the results (F1 score: 95.45%):

Class	Control	COVID-19
Control	94.12%	5.88%
COVID-19	1.2%	98.78%

Model weights are here. Train and test splits of the data are in ictcf_train.txt and ictcf_test.txt.

COVID-Single-Shot-Model Project

Presentation at the University of Maryland 09-12-2020. The content is mainly the model on github: simultaneous segmentation and COVID-19 prediction, the model is trained from scratch.

Preprint oin medRxiv:

Single-Shot Lightweight Model For The Detection of Lesions And The Prediction of COVID-19 From Chest CT Scans

BibTex:

@article {Ter-Sarkisov2020.12.01.20241786,
	author = {Ter-Sarkisov, Aram},
	title = {Single-Shot Lightweight Model For The Detection of 
	Lesions And The Prediction of COVID-19 From Chest CT Scans},
	year = {2020},
	doi = {10.1101/2020.12.01.20241786},
	publisher = {Cold Spring Harbor Laboratory Press},
	journal = {medRxiv}
}

Conceptually the model is similar to COVID-CT-Mask-Net, but there are a lot of new functionality, so I decided to create a new repository. Of the models presented in the paper, I uploaded the architecture and the weights for the one trained from scratch with two parallel branches (segmentation + classification).

Single Shot Model:

The model

Region of Interest (RoI) module with two branches

RoI Batch To Feature Vector

Download the pretrained weights into a directory called pretrained_weights. The model uses ResNet18+FPN without the last block, to reduce the number of weights. To evaluate the model on the segmentation test split:

python3 evaluation_seg_branch.py --ckpt pretrained_weights/modelA.pth --rpn_nms_th 0.75 
--roi_nms_th 0.5 --confidence_th 0.75 --device cuda --box_detections 128

You should get these results:

Model	AP@0.5	AP@0.75	mAP@[0.5:0.95:0.05]	Model size
SSM (ResNet18+FPN)	57.99%	38.28%	42.45%	8.27M

To evaluate the classification branch:

python3.5 evaluation_class_branch.py --ckpt pretrained_weights/modelA.pth --test_data path/to/test/data/ 
--device cuda  --box_nms_thresh_classifier 0.75 --box_detections 128

You should get this confusion matrix, COVID-19 sensitivity of 93.16%, F1 score of 96.76%.

	Control	CP	COVID-19
Control	9322	123	5
CP	174	7139	82
COVID-19	27	277	4042

To train from scratch, make sure you have a directory --train_seg_data_dir with --imgs_dir and --gt_dir subdirectories for the segmentation branch and --train_class_data_dir for the classification branch. The links to the data are here: http://ncov-ai.big.ac.cn/download, the train/test/validation splits are in .txt files above and in the source split: https://github.com/haydengunraj/COVIDNet-CT/blob/master/docs/dataset.md.

On a GPU with 8Gb VRAM 50 epochs should take about 5 hours.

For any questions, contact Alex Ter-Sarkisov, alex.ter-sarkisov@city.ac.uk