This model assists the radiologist to segment lesion and do positive/negative classifications based on lung CT.

Intended Use

Chest CT Dicom series of 1.25mm and 7.25mm.


The diversity of the training data is limited. The evaluation of the model is not on a comprehensive dataset.

Information on training data

The model uses deep neural networks, trained on data from several sources in China in an end-to-end manner.
30% of Training data has slice thickness of 1.25mm and the remain part is in 7.5mm. Half of the data is negative.

For positive data, pneumonia region volume/lung region volume distribution is
[0.1,0.25]: 21.4%
[0.25,0.5]: 18.3%
[0.5:0.75]: 6.0%
[0.75,100]: 1.6%

Model performance metrics

Accuracy: 0.9 (negative/positive)

Dice score on pneumonia region segmentation: 0.7

Average Dice score on lung lobe segmentation: 0.95

Model History

The model has been updated to display pneumonia quantification per lung lobe.

See more ▼less ▲

The model structure is the same as before, however, more training data labelled by experienced radiologist/with PCR results are added in.


We increase the ratio of 1mm CT dcm significantly in the Training set for this version.



This high-tech company specializes in artificial intelligence in medicine.