Physiology based unsupervised learning for segmentation of COVID-19 lesions in chest 3D CT scans.

BACKGROUND

The existing deep learning (DL) segmentation methods of COVID-19 lesions in chest CT scans require tedious manual labels for training. This is labor-intensive and time-consuming for radiologists, and also limits the application of DL methods in clinical practice.

PURPOSE

To develop an unsupervised DL segmentation method that does not need the participation of radiologists to provide manual labels for training.

METHODS

Two hundred chest 3D CT scans of 48 COVID-19 patients (including 48 baseline CT scans and 152 follow-up CT scans) were used as training set and another 65 chest 3D CT scans were used as test set. A novel self-learning segmentation method was proposed to train a DL segmentation model based on the intrinsic physiology information of the patients instead of the external professional knowledge of the radiologists. The proposed method comprised two modules: a self-annotation module and a training module. In self-annotation module, the progression of COVID-19 pneumonia and the change of lung volumes between a pair of baseline CT and follow-up CT were combined to automatically annotate the abnormal lung region (referred as self-label) in the follow-up CT. Then in training module, the follow-up CT scans and the corresponding self-labels were used to train a DL segmentation model (referred as Model-S). The COVID-19 lesions segmented by Model-S in the test set were quantitatively compared with the ground truth. To further evaluate the proposed method, the manual labels of 152 follow-up CT scans were delineated by radiologists and compared with the corresponding self-labels. Then another DL segmentation model (referred as Model-M) was trained based on these manual labels. The segmentation results between Model-S and Model-M in the test set were also compared.

RESULTS

As for evaluating the self-annotation results, the Dice similarity coefficient (DSC) between self-labels and manual labels in follow-up CT scans was 93.0% ± 10.5%. The lesion volumes of self-labels and manual labels had a high Spearman correlation (r = 0.99; p < 0.001). As for evaluating the self-learning segmentation results, the DSC between the lesions segmented by Model-S and ground truth, and between the lesions segmented by Model-M and ground truth were 82.1% ± 6.0% and 83.4% ± 5.2%, respectively. The lesion volumes segmented by Model-S and the ground truth had a high Spearman correlation (r = 0.99; p < 0.001). Besides, the DSC between the lesions segmented by Model-S and segmented by Model-M was 87.5% ± 4.0%. The lesion volumes segmented by Model-S and Model-M also had a high Spearman correlation (r = 0.98; p < 0.001).

CONCLUSIONS

The proposed unsupervised segmentation method can automatically segment COVID-19 lesions accurately in chest 3D CT scans without requiring radiologists to provide any manual labels for training in the establishment process of the method.