Unsupervised Segmentation Of COVID-19 Lesions Based On Knowledge Distillation

In recent years,Corona Virus Disease 2019(COVID-19)has spread around the world which is threatening people’s lives and health.In the fight against this epidemic,computer-aided diagnosis systems on chest CT images can play a key role in quick diagnosis and assessment of COVID-19.The important step of these systems is to automatically segment COVID-19 lesions on CT images.The accurate segmentation of COVID-19 lesions can directly display the location,size and texture of the lesions,which is helpful for disease analysis and doctors’ diagnosis and treatment.At present,most of the COVID-19 lesion segmentation work is developed based on supervised methods and require large-scale annotated data that is difficult to obtain,while methods based on unsupervised anomaly detection can segment lesions only using easily available normal data(data from healthy lungs).An existing unsupervised anomaly detection method for COVID-19 lesion segmentation is proposed based on the reconstruction error framework.Due to the limitation of the framework itself,this method is very dependent on the quality of reconstructed image,and it is not suitable for high-resolution images,overall performance is poor.Therefore,this paper investigates and improves a knowledge distillation framework,which uses feature discrepancy of the teacher network and the student network to locate anomalies,instead of the reconstruction error framework that uses image discrepancy of original image and reconstructed image to locate anomalies.Specifically,this paper proposes a pixel-level and affinity-level knowledge distillation algorithm which suitable for unsupervised segmentation of COVID-19 lesions on CT images.The algorithm obtains a pre-trained teacher network with rich semantic knowledge of CT images by constructing and training an auto-encoder at first,and then trains a student network with the same architecture as the teacher by distilling the teacher’s knowledge only from normal CT images,and finally segment lesions using the feature discrepancy between the teacher and the student networks on lesion images.Aiming at the problem that insufficient knowledge distillation will affect the distribution of normal images learned by the student network and thus affect the segmentation accuracy,this method not only uses the traditional pixel-level distillation,but also designs the affinity-level distillation that takes into account the relationship between pixels,so as to fully distill effective knowledge.To verify the effectiveness of the proposed pixel-level and affinity-level knowledge distillation algorithm,this paper evaluates on one private dataset and two public datasets with three evaluation metrics.Experiments show that compared with other existing unsupervised anomaly detection methods,the best AUROC index of the three datasets is 2.31%,1.54% and2.69% higher respectively,AUPRC index is 14.84%,4.51% and 15.14% higher respectively,and DSC index is 18.23%,9.6% and 14.9% higher respectively.It can be seen that the algorithm in this paper has greatly improved the performance of unsupervised segmentation of lesions.