At current stage, the COVID-19 epidemic has affected the global social, economic and livelihood development. The prevention and control work of COVID-19 epidemic is very important. "Early detection, early report, early isolation, early diagnosis and early treatment" is the key to the prevention and treatment of COVID-19. As a routine examination of pulmonary lesions, CT plays an important clinical role in the early diagnosis and dynamic assessment of COVID-19[7].
CT image is a special 3D image, so its focus detection task also requires the output of 3D external compact box. For 3D data, 3D convolution has more advantages over traditional 2D convolution[8]. The convolution kernel of 3D convolution can be traversed in 3D space to obtain 3D spatial features, and it has a stronger ability to represent lesions, whereas 2D convolution can only extract features in a plane. This system takes 3D convolution as the basic unit when building the model. While improving representational ability, 3D convolution also raises the amount of computation. In order to ensure the training efficiency of the model and the real-time performance of the application, the number of channels in 3D convolution layer was reduced and the step size of the spatial pooling layer was increased to meet the requirement of efficiency when designing the model. The recognition and segmentation of COVID-19 lesions can be achieved through the deep learning of AI. The entire sequence of a CT examination was put into the detection and segmentation models respectively to obtain the pneumonia lesion detection box and the pneumonia lesion segmentation mask. The two sets of results were then fused to obtain the final detection results. On the basis of the detection and segmentation of lesions, the volume and average density of lesions can be obtained through cumulative calculation based on DICOM information, including pixel spacing, layer thickness and layer spacing.
During the operation of the system, the diagnostic physician can directly locate the lesion at the image level after selecting the lesion in the list of screening results. Based on the location and attributes of the lesions, the system can generate a report of its imaging manifestations. When the scope of the pneumonia lesions exceeds a specific threshold of the total lung volume, a warning of "high proportion of abnormalities and suspected pneumonia" will be issued. The diagnostic physician can also remove the false positive lesions in the results, and give a more realistic lesion size and volume.
During the study, the AI image-assisted diagnosis system performed 97.78% sensitivity and 92.31% specificity for the diagnostic efficiency of COVID-19. Its accuracy rate was up to 97.09%, proving its application value in the early diagnosis of COVID-19. In the chest imaging analysis of COVID-19, it was found that most lesions were multiple ground glass opacity in both lungs. Some lesions showed consolidation and fibrosis in the course of disease, which was consistent with the report of Salehi S et al[9].
Due to the wide distribution and rapid progression of most pneumonia lesions, the detection rate of the system for a single lesion was only 89.89%, and the false alarm rate was 27.97%. The 64 missed diagnosis markers were all small ground glass opacity. Although these lesions were not marked in the pneumonia list, they appeared in the nodule screening function list. This is related to the algorithm model and parameter setting of the system which can distinguish between nodules and pneumonia by the size of the lesions. With the continuous development of deep learning increasing clinical application data, the AI image-assisted diagnosis system will be continuously improved and upgraded[10].
Pulmonary vascular shadow and respiratory movement artifact accounted for more than 80% of the 221 misdiagnosed markers. This is related to the quality of the image data, which is affected by the parameter setting of image equipment and algorithm reconstruction. In this study, the slice thickness of most CT scans was 3mm, and the scanning time was about 45s. The breathing movement due to the patient's lack of breath control affected the image quality, thus affecting the recognition efficiency. Some studies suggested that thin CT slice will contribute to lesion recognition efficiency. The recommended thickness is 1.25mm[11]. Further clinical trials should be conducted to verify its specific impact.
Through comparing the CT reexamination of 24 patients, it is found that under the condition that all patients showed recovery during hospitalization, the lesions in the last CT scan all showed different degrees of absorption compared with the first scan, which is demonstrated in the reduction in lesion size and lesion percentage of total lung volume and the decrease in mean lesion density. The AI image-assisted diagnosis system is able to present the dynamic changes of the patient's CT results with higher accuracy and shorter response time. Therefore, image diagnostician and clinicians can efficiently assess the clinical efficacy of treatment, and provide standards as well as basis for disease evaluation, isolation removal and discharge. At the same time, the system can quickly process massive data and make rapid and accurate comparison among multiple reviews, which can reduce the workload of doctors and avoid the subjective bias of doctors.
In addition, 7 patients in this study were negative on chest CT. Three of them underwent two chest CT examinations. However, all the patients were tested positive by fluorescence RT-PCR with nucleonuclear acid of novel coronavirus. This indicates that negative chest CT results are not the absolute standard to exclude COVID-19. Therefore, suspected patients with negative CT results should be treated with caution in clinical practice to prevent missed diagnosis. The differential diagnosis of COVID-19 with other viral pneumonia, bacterial pneumonia and mycoplasma pneumonia is also a key and difficult area in clinical work, which will also be a new direction of AI image-assisted diagnosis system.