Exploring the Value of Lung Texture Features in Distinguishing Usual and Non-specic Interstitial Pneumonia

Backgground At present, the most common types of interstitial pneumonia are usual interstitial pneumonia (UIP) and non-specic interstitial pneumonia (NSIP), and different types have different prognosis. In addition, if there is a mixture of different classications, it will be dicult for radiologists to diagnose, and it will make clinical treatment dicult. Therefore, clinicians urgently need new imaging methods to solve such problems. This article aims to explore the CT lung texture images of UIP and NSIP to provide evidence for the identication of UIP and NSIP. Methods A retrospective analysis of 96 cases of interstitial pneumonia diagnosed by the Department of Pathology and the Aliated Hospital of Guilin Medical College. Among them, there are 40 cases of UIP and 56 cases of NSIP. All patients are scanned by CT. Lung Intelligence Kit was utilized to perform lung segmentation and texture feature extraction. Variance analysis, least absolute shrinkage and selection operator (Lasso) and multivariate logistic regression were used to select effective features. Finally, a multivariate logistic regression model was constructed to identify two kinds of interstitial pneumonia. Receiver operating characteristic (ROC) curve, area under the curve (AUC), sensitivity, specicity were used to evaluate performance of the constructed model. We used the LK software to segment the two sets of lungs. Feature calculation and selection were performed on the data of the two groups of interstitial pneumonia after lung segmentation, the logistic regression model was established for the selected features, and the ROC curve was drawn. Results A total of 100 texture features are extracted from the whole lung segmented by LK, and nally 8 features are left after feature reduction. The above-mentioned values of UIP and NSIP of the training group are greater than those of the test group. Conclusions It is possible to distinguish UIP and NSIP by using LK software to extract lung texture in CT images.


Introduction
Interstitial pneumonia is a lung disease characterized by pulmonary interstitial in ammation and progressive pulmonary brosis. It has many causes and different types have different prognosis [1] .
Currently, the most common types of interstitial pneumonia are usual interstitial pneumonia (UIP) and non-speci c interstitial pneumonia (NSIP). However, the two types of interstitial pneumonia are di cult to distinguish on the initial clinical manifestations or CT images, especially when they are atypical. In addition, the treatments for the two types of interstitial pneumonia are different. Misdiagnosis and wrong treatment of patients may lead to unforeseen complications [2] .
According to related literature reports, pathological biopsy has good performance in the classi cation of interstitial pneumonia, but it will cause a series of complications and reduce the survival rate of patients [3] . Therefore, it makes sense to nd a non-invasive way to distinguish between the two types of interstitial pneumonia. In recent years, radiomics research as an emerging research eld has developed rapidly. It extracts high-level and unrecognizable features from the images; and has shown potential in variable tumor or non-tumor diseases [4][5][6] . A large number of radiomics studies have been published in lung diseases. However, so far, no studies have been found to distinguish interstitial pneumonia types by radiomics.
In this study, we extracted the double-lung texture features of patients with interstitial pneumonia based on Lung Intelligence Kit (LK; GE Healthcare), and performed feature selection and model construction, aiming to provide imaging evidence for the accurate diagnosis of UIP and NSIP. In this work, our contributions are as follows: (1) Using LK software to perform reliable lung segmentation and texture features extraction. (2) Constructing a radiomics model to support UIP or NSIP diagnosis.

Patients
In this study, the data of patients with interstitial pneumonia who were treated in the A liated Hospital of Guilin Medical University from March 2015 to July 2020 were retrospectively collected. Whether the patient has interstitial pneumonia is usually determined by the results of frozen lung biopsy during the treatment process or HRCT examination at during the rst visit.
The patient's alveolar lavage, lung function, blood oxygen saturation, etc. were consistent with UIP or NSIP respectively, and preliminary classi cation was performed. The patient continued to recheck after treatment in HRCT, the assessment is divided into effective and ineffective groups, the effective group is NSIP, and the ineffective group is UIP. Finally, inter-division communication is conducted with the clinic, and the clinical diagnosis is UIP or NSIP, and the corresponding grouping is performed. Patient selection criteria: 1. All patients with interstitial pneumonia 2. The lesion is obvious and clear on HRCT.
Exclusion criteria for patients: 1. There is no pathological result and no clinical diagnosis type. 2.
Pathologically and clinically diagnosed non-UIP and non-NSIP patients.
During this period, a total of 96 patients with interstitial pneumonia were collected, of which 40 were in the UIP group (25 males and 15 females), aged 42-91 years old, with a median age of 67 years and an average age of 66.65 years; 56 in the NSIP group Cases (31 males and 25 females), aged 35-89 years old, with a median age of 58 years and an average age of 60.87 years.

HRCT scanning
CT scans were performed in all cases with GE light VCT 64, GE Optima600 128 or GE Revolution 256-slice CT scanner. The scanning range is from the apex of the lungs to the line 2 cm below the parallel line of the double costal phrenic angle. Scanning parameters are as follows:128KV, automatic mA, noise index 8, layer thickness and layer spacing of 5mm. The image under secondary thinning and reorganization, layer spacing and layer spacing are both 1.25mm after reorganization.

Image analysis
Two associate senior or higher physicians who do not understand clinical and pathological classi cation will make the diagnosis of interstitial pneumonia type on all patients selected for the study. If two doctors make the same type of diagnosis and clinical diagnosis, the diagnosis is still If they are consistent, they are classi ed as UIP or NSIP accordingly; if the two doctors' diagnosis on the image is inconsistent, they need to communicate again and communicate with the clinic to make the nal diagnosis. They are also divided into UIP or NSIP group.

Lung segmentation and feature extraction
In this study, lung segmentation and feature extraction were conducted by LK software (Lung Intelligence Kit, GE Healthcare). LK is a lung image analysis software, which can automatically segment lung lobe based on DenseVNet, and extract texture features through the integrated pyradiomics package. The schematic diagram of the segmentation result is shown in Fig. 1. Each patient extracts 100 features in 6 categories, including: histogram features, morphological parameters, gray level symbiosis Matrix features, gray-scale dependent matrix, gray-scale run-length matrix, gray-scale area matrix.

Feature selection and model construction
The data set is randomly divided into training group and test group at a ratio of 7:3. All cases in the training group are used for model training, and the test group is used to independently evaluate the performance of the model. Replace the missing values and outliers in the data with the mean; and use the Z-score method to standardize and preprocess the data. The features with a variance of 0 are removed; after that, we use the LASSO (Least absolute shrinkage and selection operator) method to reduce dimensionality, and nally build a multiple logistic regression model based on the selected features, and draw Nomogram.

Statistical analysis
The receiver operating characteristic (ROC) curve was used to evaluate the performance of the multiple logistic regression model. Related accuracy, sensitivity, speci city and area under the curve (AUC) were calculated. All statistical analyses in this study were performed using R language (Version 3.5.1, https://www.r-project.org). Two-tailed p value < 0.05 indicates statistical signi cance.

Model building results
In the training group and the test group, the logistic regression model was used to evaluate the accuracy, those of the test group, as shown in Fig. 3(A-B) and Table 2 .

Discussion
At present, the diagnosis and classi cation of interstitial pneumonia mainly rely on HRCT, and the treatment methods used for different classi cations are not all the same. On the other hand, the in ammation and brosis of interstitial pneumonia may have the same performance on HRCT [8][9] .For atypical interstitial pneumonia, HRCT seems powerless, so it usually requires pathological biopsy for further examination. However, according to reports in the literature, pathological biopsy will cause a series of complications and reduce the survival rate of patients. At the same time, pathological biopsy can only represent the structure of local lung lesions, and cannot fully understand the condition of lung lesions. Therefore, using deep learning for automatic lung segmentation and obtaining high-level texture parameters of the lungs can effectively avoid doctors' subjective interference factors, obtain more comprehensive information about interstitial pneumonia, and achieve more accurate diagnosis.
Pang et al. used imaging omics to segment the HRCT of ILD patients and found that it has higher accuracy and overall performance than conventional methods. The segmentation results prove the necessity of denoising and the practicality of radiological features for segmentation [9] .From this, we think of using radiomics to explore the ability to distinguish between UIP and NSIP. Through this study, for the rst time, we used LK software to segment the lungs of all patients with interstitial pneumonia, and extracted lung texture features, and established logistic regression to distinguish UIP from NSIP based on these high-level features. Through the ROC curve of the logistic regression model, it can be seen that the classi cation model of radiomics shows high diagnostic e ciency,and the AUC values of the training group and the test group are 0.952 and 0.838, respectively.
The F1 value is a new machine learning indicator. The F1 value is the weighted average of precision and recall. The best F1 value is 1 and the worst is 0. In this study, the F1 values of the training group and the test group were 0.868 and 0.696, respectively, which showed the good diagnostic performance of this model.
The remaining eight features in this study re ect the gray value distribution, texture feature and spatial distribution of UIP and NSIP patient images, respectively. Among them, "original_gldm_DependenceVariance", "original_gldm_DependenceNonUniformity", and "original_shape_Maximum2DDiameterColumn" re ect the apparent texture thickness, shape, gray scale information of the lung. "Original_glrlm_GrayLevelVariance" re ects the variance of Gray Level in gray level co-occurrence matrix. "Original_glszm_ZonePercentage", "original_glszm_SizeZoneNonUniformityNormalized", "original_glszm_GrayLevelNonUniformityNormalized", and "original_glcm_DifferenceVariance" re ect the complexity of the lesion site, the degree of change in the level of complexity, and the level of change. These indicate that the texture characteristics of the lungs are different in UIP and NSIP patients.
There are some limitations in this study: Firstly, it is a single-center, small-sample study. The results should be veri ed by multi-center large sample data. Secondly, the grouping of UIP and NSIP is partly based on clinical diagnosis. This method should be feasible, because the clinical manifestations of UIP and NSIP in the middle and late stages are quite different. Thirdly, a machine learning model was established to distinguish two types in interstitial pneumonia, and deep learning algorithms may provide better discrimination performance, which will be researched in the future.Fourthly, the CT images obtained in different devices may affect the image analysis.Fifthly, Most of the patients with interstitial lung disease are elderly and accompanying pulmonary edema, infection, or emphysema can be seen. I think these ndings may affect the results.
In general, this article investigates the potential of lung texture to distinguish between UIP and NSIP, which can bring bene ts to the diagnosis and treatment of patients with interstitial pneumonia Abbreviations UIP usual interstitial pneumonia The Institutional Review Board of Guilin medical university Hospital approved this retrospective study and waived the requirement for written informed consent due to its retrospective nature.All methods were performed in accordance with the relevant guidelines and regulations.

Consent for publication
Not applicable.

Availability of data and materials
The datasets analyzed in this study are available from the corresponding author on request.