CT features and artificial intelligence quantitative analysis of recovered COVID-19 patients with negative RT-PCR and clinical symptoms CURRENT STATUS: POSTED

Objectives: To evaluate imaging features and performed quantitative analysis for mild novel coronavirus pneumonia (COVID-19) cases ready for discharge. Methods: CT images of 125 patients (16-67 years, 63 males) recovering from COVID-19 were examined. We defined the double-negative period (DNp) as the period between the sampling days of two consecutive negative RT-PCR and three days thereafter. Lesion demonstrations and distributions on CT in DNp (CT DN ) were evaluated by radiologists and artificial intelligence (AI) software. Major lesion transformations and the involvement range for patients with follow-up CT were analyzed. Results: Twenty (16.0%) patients exhibited normal CT DN ; abnormal CT DN for 105 indicated ground-glass opacity (GGO) (99/125, 79.2%) and fibrosis (56/125, 44.8%) as the most frequent CT findings. Bilateral-lung involvement with mixed or random distribution was most common for GGO on CT DN . Fibrous lesions often affected both lungs, tending to distribute on the subpleura. Follow-up CT showed lesion improvement manifesting as GGO thinning (40/40, 100%), fibrosis reduction (17/26, 65.4%), and consolidation fading (9/11, 81.8%), with or without range reduction. AI analysis showed the highest proportions for right lower lobe involvement (volume, 12.01±35.87cm 3 ; percentage; 1.45±4.58%) and CT-value ranging –570 to –470 HU (volume, 2.93±7.04cm 3 ; percentage, 5.28±6.47%). Among cases with follow-up CT, most of lung lobes and CT-value ranges displayed a significant reduction after DNp. Conclusions: The main CT imaging manifestations were GGO and fibrosis in DNp, which weakened with or without volume reduction. AI analysis results were consistent with imaging features and changes, possibly serving as an objective indicator for disease monitoring and discharge.


Introduction
The ongoing pneumonia caused by a novel coronavirus (COVID-19) has been widely studied by scholars and people worldwide. It was first reported as an "unknown viral pneumonia" with a limited number in one city of China in December 2019 [1][2][3]. In the following months, infection spread rapidly across China and many other countries around the world [4][5][6]. On January 30, 2020, the World Health Organization (WHO) declared a Public Health Emergency of International Concern (PHEIC) [7]. Chinese public health agencies promptly invoked measures to contain the outbreak in China. Moreover, experts from different fields of study are working together to improve the recognition, diagnosis and treatment of patients. At present, many patients in China with mild disease recover.
Chest CT has been proven to be important for the diagnosis and monitoring of the disease course of COVID-19 [8][9][10]. Many studies have reported typical CT features, including ground-glass opacity (GGO), consolidation, and/or interstitial changes with a peripheral or bronchovascular bundle distribution [11][12][13][14][15][16][17]. Nonetheless, studies on changes in CT features with the course of the disease are relatively limited [14][15][16]18]. Pan. et al [14] summarized four stages of the disease and found that the lesions observed on imaging gradually disappeared after two weeks. In general, more information about the imaging characteristics of patients who have recovered and are ready for discharge is needed.
Here, we describe the CT features of mild cases ready for discharge, with two negative reverse transcription-polymerase chain reaction (RT-PCR) results and no symptoms for at least 5 days, and attempted to perform a quantitative analysis of the findings. We hope that a better understanding and quantitative analysis of the CT features of recovered patients will help to more accurately judge discharge.

Materials And Methods
The study was approved by the ethics committee of our hospital. Informed consent was waived due to the nature of the retrospective study.

Study Population
From February 10, 2020, to March 10, 2020, CT images for 135 patients recovering from COVID-19 at the mobile cabin hospital in Hongshan Gymnasium, Wuhan, Hubei, were reviewed retrospectively. We defined the double-negative period (DNp) as the time period between the sampling days of two consecutive negative RT-PCR detections and three days thereafter. The inclusion criteria were 1) patients with mild disease recovering from COVID-19 with negative RT-PCR for two consecutive lower respiratory specimens (sampling time interval of at least 24 hours); 2) no abnormal clinical symptoms and laboratory indicators for at least 5 days; 3) at least one chest CT scanning during DNp. Ten patients were excluded due to unavailability of CT images during DNp (n = 1) and persistent RT-PCR positivity during the study period (n = 9).

CT Protocol
CT scans for all patients were performed using a mobile CT Unit (CT-NeuVz Prime, NEUSOFT) with a single inspiratory period. The scan parameters included the following: tube voltage of 120 kVp, tube current of 100-200 mA, detector collimation of 64 or 128× 0.625 mm, field of view of 350 mm×350 mm, and matrix size of 512 × 512. All images were reconstructed using a medium sharp reconstruction algorithm with a slice thickness of 5 mm.

Chest CT evaluation in DNp
The CT images during DNp (CT DN ) and follow-up CT after CT DN for all patients were transmitted to the remote image diagnosis cloud platform through the 5G network. Two radiologists (Reader 1 and Reader 2 with 5 and 15 years of experience in chest CT diagnosis, respectively) reviewed the CT images slice by slice. CT demonstrations and lesion distribution were recorded. When inconsistency occurred, the another radiologist (Reader 3) with 30 years of chest CT diagnosis experience would make the decision.
Comparative evaluation of follow-up CT For patients with available follow-up CT images after CT DN , the first follow-up CT was additionally collected for comparison with CT DN . The same two radiologists assessed transformations in major CT demonstrations and lung involvement ranges. The former were divided into two categories: improvement and unchanged. Transformation types of major CT demonstrations and ranges of involvement lungs were further analyzed in the improvement group, and transformation of involvement areas was defined as decreased or unchanged.

Quantitative AI analysis
An AI-assisted screening and epidemic intelligence monitoring system for COVID-19 pneumonia,

Patient characteristics
Sixty-three men and 62 women (average age 48±13, range 16-67) were included in the study. The period between the onset of initial symptoms and CT DN was 22±10 days. T DN was significantly longer for cases with the the presence of lesions (p = 0.007), GGO (p = 0.001) and consolidation (p = 0.035).
Forty-four of 125 (35.2%) patients underwent follow-up CT after CT DN (Table 1). There were 99, 68 and 18 patients presenting GGO, fibrous lesions and consolidation, respectively (Table 3). Bilateral-lung involvement with a mixed and random distribution was the most common for GGO in DNp (55/99, 55.6%). Fibrous lesions were also more likely to affect both lungs (53/68, 77.9%), and it was basically flat for only subpleural distribution and mixed/random distribution (24/68, 35.3% vs. 29/68, 42.6%). The highest distribution proportion under the subpleural area of the right lung was found for consolidation (10/18, 55.6%), and mixed/random and only bronchovascular distributions were rare.

Transformations in CT manifestations
The counts of CT demonstration transformations and the range of involvement lungs are presented in  Table 1). For 66/125 (52.8%) patients with AI results of 0, 46/66 (69.7%) were observed by the radiologists to have small and thin GGOs on CT images, whereas AI did not.
Among 44/125 (35.2%) cases with follow-up CT, the involved volume and percentage decreased significantly (Z = -3.55, -3.49; p <0.001). Figure 4 shows the AI analysis results (blue line) of a typical pattern of GGO decreasing and thinning on follow-up CT. The lesion volume and its percentage of the right upper and lower lobe, and left upper lobe were significantly reduced on follow-up CT after CT DN (Table 5). For different CT-value ranges, a significant reduction after DNp appeared in all ranges of CT values expect for the volume percentage between 30HU and 60HU. For 33 cases with lesions detected by AI from 44 cases with follow-up CT, the same changes were found as those in the 44 cases described above (Supplementary Table 2).

Discussion
We summarized the imaging characteristics and performed AI quantitative analysis for COVID-19 patients with mild disease and ready for discharge. During DNp, the main imaging manifestations were GGO and fibrosis with reticulation inside. On the follow-up CT, GGO and fibrosis further weakened, with or without volume reduction. Enlarged MLN was observed. The AI quantitative analysis results were consistent with the imaging features and the changes. Thus, AI may serve as an indicator for disease monitoring and discharge. However, the diagnostic ability of AI in the initial diagnosis of COVID-19 deserves further study due to insufficient detection for GGO.
In contrast to the early stage of COVID-19, GGO during DNp is rarely accompanied by other signs, such as "crazy paving" pattern [14]. The GGO density in DNp was thinner and more homogenous.
According to the latest guidelines of Diagnosis and Treatment of Pneumonitis Caused by COVID-19 (trial seventh version) [21], one of the discharge criteria is acute exudative lesions that are significantly improved on chest CT. We need to understand that GGO does not always represent exudative lesions. Pan. et al found that consolidation gradually disappeared with GGO and subpleural parenchymal bands remaining after two weeks [14]. This was confirmed in the 44 follow-up cases in this study. The results showed that the original GGO was further reduced, diluted or even disappeared; In addition, the fibrosis on CT DN changed to GGO with different sizes after follow-up. The thinner and more homogenous GGO in DNp might be a remnant of GGO, consolidation or fibrosis after absorption and does not necessarily affect the discharge assessment. Careful observation and combination with comparative observation are needed for a final judgment.
Fibrosis was the second most common imaging manifestation in DNp. Fibrosis is characterized by heterogeneous density and contraction [20]. Fibrosis may be a change from consolidation [22], suggesting the absorption of consolidation in mild cases. Fibrinous exudation in the alveolar lumen has been reported in recent COVID-19 pathological studies [23], which might be the pathological basis for fibrosis on imaging. Fibrosis can be further absorbed and change to GGO [22], which was also observed in our 44 follow-up cases. As a marked feature, reticulation is widely observed in fibrosis. Different from the "crazy paving" in GGO at the early stage, the reticulation in the fibrosis is more rigid and thicker. Regardless, the pathological mechanism underlying rigid reticulation in fibrosis is not clear. It is possible that fibrinous exudation in the interlobular septal may occur.
Alternatively, the pathological changes of the interlobular septum were slower than those of the alveolar lumen during the process of absorption. Obviously, these reticulated structures may also continue to be absorbed until they disappear. Overall, the range and severity of fibrosis and reticulation should be evaluated and followed-up for the determination of discharge.
Consolidation with homogenous density and air bronchogram was not common during DNp. In this study, CT DN was performed at 22 days after the onset of the initial symptoms. Consolidation was reported to reach its peak at approximately 10 days after the initial symptoms and gradually absorbed after 14 days [11]. At the follow-up in this study, part of the original consolidation was converted into fibrosis, and part of it was converted to GGO. The absence or small consolidation may be a sign of DNp.
Enlarged MLN has not commonly been observed in the early stage [11][12][13][14][15][16][17]. Xiong. et al [24] found an increased number of enlarged MLN during follow-up in the course of the disease. On CT DN , enlarged MLN was observed in 45(36%) patients, especially with a size of 1.0-1.5cm. It was difficult to judge when MLN enlargement occurred without imaging data of the entire course. MLN involvement might occur later than lung involvement and also change at a slower rate. In general, enlarged MLN might not affect discharge determination.
Fibrosis stripes were slight during DNp in mild patients, suggesting a good prognosis. Other common signs in early images were rarely observed in DNp, which is suggested that these signs are mainly related to early pathological changes. After follow-up, the changes in fibrosis stripes and other rare signs on CT DN were not obvious. The absence of such signs may be a criterion for discharge.
The data obtained from AI quantitative analysis were consistent with the imaging features during DNp. The average volume of the lesion was small, accounting for only 1.21 ± 1.81% of the entire lung. GGO was the dominant lesion, with -570 to -470 HU accounting for the highest proportion. be carried out to provide more data to support AI quantitative assessment of pathological processes and even outcome prediction.
Overall, AI was not precise enough to detect GGO with very low density (shown as yellow arrow in Figure 4). In 46 (69.7%) patients with AI results of 0, small and thin GGOs on the image were observed by the radiologists. This might be related to the density interval of the observation setting of the software as a lower density GGO less than -570 HU cannot be detected. As we observed, the thin and homogenous GGO in DNp transformed from consolidation or fibrosis during absorption. As the presence or size of a GGO does not affect the determination of discharge, the AI result of 0 might still serve as a criterion for discharge in clinical work, even though some thin GGOs may still exist.
There are still some limitations. First, it was a single-center study and the number of cases was also limited. Second, as a retrospective study, only patients with mild disease were included without severe disease. Additionally, all patients were transferred from other hospitals to the mobile cabin hospital. Early digital images at the onset or diagnosis were lacking, making it difficult to compare.
Finally, with a scanning thickness of 5 mm, some details may not be as good as those obtained with thin slices. However, the scanning protocol was in accordance with the work flow of the mobile cabin hospital, with multiple CT examinations for each patient and relatively limited medical staff.
In conclusion, the main CT imaging manifestations were GGO and fibrosis in DNp. On the follow-up CT, GGO and fibrosis will further weaked with or without volume reduction. AI analysis results were consistent with imaging features and changes, possibly serving as an objective indicator for disease monitoring and discharge.   RT-PCR: reverse transcription-polymerase chain reaction T DN : the period between the onset of initial symptoms and the CT DN    The details of different CT findings in DNp. a: Multiple thin and homogenous ground-glass opacities (GGOs) (box) and the pulmonary nodule (arrow). b: Thickened small blood vessels traveling inside the thin GGO. c: Small local consolidation and air bronchogram (box) in the right lung and fibrosis (arrow) in the left lung. d: Reticulation (arrow) (more rigid and thicker than "crazy-paving" reticulation) in the subplural regions of both lungs, which was widely observed in fibrosis. e: Fibrosis stripe (arrow). f: Thickened bronchial wall (arrow). g-i: Mediastinal lymph node size <10 mm (g), 10-15 mm (h) and >15 mm (i) (arrow). DNp: double-negative period (the period between the sampling days of two consecutive negative RT-PCR detections and three days thereafter).

Figure 2
Counts of CT demonstration transformations and range of involvement lungs in 41 cases with overall CT improvement. For each lesion type, the column bar on the left represents the count of cases with the corresponding lesion types; the column bar on the right represents the change in the lesion type and its count (refer to the legend on the right).
GGO: ground-glass opacity.  The AI analysis results of a typical pattern of GGO decreasing and thinning on follow-up CT.
GGOs on right lower lobe recognized by AI (blue line) significantly decrease on follow-up CT compared with CTDN, though AI is not precise enough to detect GGO with very low density (yellow arrow, thin but visible GGOs are not recognized by AI