We identified the value of CT radiomic features for predicting RT-PCR negativity and established a predictive model based on CT radiomic features combined with clinical data in COVID-19 patients during the recovery period. With AUCs of 0.811 and 0.812 for the training and testing datasets, respectively, we expect the model to help doctors effectively predict RT-PCR negativity during clinical treatment.
The unsatisfactory sensitivity of RT-PCR detection is a major concern [2, 3, 11, 12]. To avoid the possibility of false negative RT-PCR in our study, we included patients with repeated RT-PCR tests (average times: 6; range 3–12) during hospitalization. Only the patients with consistent results of the consecutive RT-PCR tests were included to ensure true negative or positive RT-PCR status for the corresponding CT. A 2-week isolation after discharge was further performed to avoid any possibility of false negative RT-PCR.
Accurate lesion segmentation is the key to feature extraction and model construction. We used the semiautomatic pneumonia segmentation software based on a deep learning algorithm combined with manual adjustment. Recently, deep learning techniques have been widely used in the detection of COVID-19 lesions in chest CT images [7–10]. Most studies [7–9] applied it to CT images in the early stage of the disease course for diagnosis and differential diagnosis, while there are few studies regarding CT images of COVID-19 patients during the recovery period. Unlike Colombi et al.’s study [13], in which the lung lobes were divided into three equal parts, our study, based on the actual segmentation of the lung lobes, detected the respiratory tract (trachea and bronchus) and lung lesions and obtained comprehensive quantitative parameters and radiomic features of lung infection for model construction.
Radiomic features played important roles in the model. Among the 10 parameters in the model, 9 were CT radiomic features. The top five radiomic features are original_firstorder_Minimum, original_gldm_Small Dependence Low Gray Level Emphasis, original_glszm_Large Area High Gray Level Emphasis, original_firstorder_10Percentile, and original_shape_Sphericity (Table 2). These indicators represent lesion internal heterogeneity of morphology, density, texture and distribution, thus indicating disease severity. The time interval from symptom onset was the only clinical parameter selected in the model, with the strongest correlation with the RT-PCR-negative group (OR = 2.84). As expected, the longer the disease course, the more patients received negative RT-PCR.
We also analyzed the lung CT quantitative parameters, but none of them were included in the model. Increased numbers, extents, and densities of ground-glass opacities (GGOs) [14] and consolidations [15] represent progression in COVID-19 patients, as well as the transformation of consolidation from GGOs [16]. Decreased sizes, extents, and degrees of such lesions could indicate improvement [15, 17–20]. In our study, the recovering patients who had a negative RT-PCR result were expected to show smaller lesion volumes and lower CT values, but the quantitative parameters were not precise enough for the changes. The high-throughput and high-dimensional radiomic features could reflect more detailed changes inside the lesions than the CT quantitative parameters.
No laboratory indicators were included in the model. Neutrophils and lymphocytes are the main hematological indicators reflecting systematic inflammation. Lymphocytopenia occurred in more than 80% of critically ill patients [21], while in an almost mild study population, only 35% of patients had mild lymphocytopenia [22]. Elevated baseline neutrophils in mild cases were not common, and only 6.3% of non-severe patients showed increases in Zhang et al.’s study [23]. Neutrophils also did not increase over the disease course for patients with mild disease and survivors [17, 24]. The patients included in our study were mild COVID-19 patients from Fangcang Shelter Hospital. Most laboratory indicators were normal or slightly exceeded normal limits, and we did not find a significant difference in lymphocytes and neutrophils between the RT-PCR-negative and RT-PCR-positive groups.
This study has several limitations. First, the data for this study originated from a single center, studies involving more research institutions are required for validation. Second, as a retrospective study, our study only involved mild COVID-19 cases and lacked severe cases. Moreover, we only built one model type and lacked comparative analysis with other model types, including decision trees, random forests and support vector machines. Finally, we did not explain the biological interpretation of the radiomic features. We are fully aware of the need for further exploration of these conclusions in subsequent studies.