A CT-Based Radiomics Nomogram To Predict The Biological Activity of Hepatic Alveolar Echinococcosis

Background This study aims to establish a computed tomography (CT) - based radiomics nomogram to predict the biological activity of hepatic alveolar echinococcosis (HAE). Methods A total of 174 HAE patients (139 for training, 35 for test) were enrolled whose CT and positron emission tomography-computed tomography (PET/CT) examinations were performed before surgery, and the biological activity was evaluated according to the PET/CT. Radiomic features were extracted from CT images, based on which radiomic scores (Rad-score) were calculated with the least absolute shrinkage and selection operator logistic regression. Three radiomics models (K-Nearest Neighbors, Logical regression, and Multilayer Perceptron), including only radiomic features and a radiomics nomogram, comprised of demographics, clinical indexes, and radiomic features were constructed respectively to predict the biological activity of HAE. The model performance was evaluated by area under curve (AUC), decision curve, and calibration curve. 30 features in total were selected as optimal radiomic features and considered as input to calculate the Rad-score. There were no signicant differences in the predictive ecacy between the combined models and the radiomics models from the perspective of the decision curve. The radiomics models was unparalleled, with an AUC of 0.952 (95%CI=0.902~0.981, P<0.0001) and 0.800 (95%CI=0.631~0.916, P<0.0020) in the training and testing cohort, respectively. The radiomics nomogram model showed great potential in identifying HAE biological activity. curve; ABZ: benzimidazoles-albendazole; HCC: SUV: Standardized uptake value; VOIs:Volumes of interest; IBSI:Standardisation Initiative; GLCM: gray level co-occurrence matrix; GLSZM: gray level size zone matrix; GLRLM: gray level run length matrix; NGTDM: neighboring gray-tone difference matrix; GLDM: gray level dependence matrix; LASSO: least absolute shrinkage and selection operator; KNN: K-Nearest Neighbors; LR: Logical regression: MLP: Multilayer Perceptron; CIs: Condence intervals; DCA: Decision curve analysis.


Introduction
Hepatic alveolar echinococcosis (HAE) is caused by the parasitic metacestode Echinococcus multilocularis, which was rare but life-threatening. Most patients are diagnosed at the advanced stage because of the hidden early symptoms of HAE [1] and can barely bene t from the radical operation [2,3], which leads patients had to complete lifelong pharmacological treatment with benzimidazolesalbendazole (ABZ) or mebendazole (MBZ) [3,4], and thorough follow-ups. Patients have to suffer from drug side effects [5] and the high cost of treatment [6]. Unfortunately, there is no an effective curative effect evaluation standard for parasiticidal drug targeting metacestode stages of the parasite [7]. 18F uorodeoxyglucose (18F-FDG) positron emission computed tomography (PET/CT), which is currently considered reliable, has been used for this purpose [8,9,10,11]. Recent studies have shown that the inactivity of PET/CT is the main drug cessation indicator [12,13]. However, complicated equipments and high healthcare costs limit its widespread acceptance as routine HAE evaluation. Thus, an economical and practical alternative is still badly needed.
As well known, characteristics have been set up to describe HAE on many imaging modalities such as ultrasound, CT, or MR images.Some feasible results in terms of lesion activity evaluation have been achieved in some studies [14,15,16,17]compared to PET/CT based on morphology and imaging features (calci cation and microcysts sign). But it remains challenging for radiologists to nd wellvalidated imaging markers to determine metacestode viability in HAE. The concept of radiomics has attracted increased attention in recent years [18]. In liver diseases, radiomics models have been involved in brosis staging, portal hypertension evaluating, and focal lesions qualitative diagnosis [19], and radiomics models constructed by incorporating clinical features of hepatocellular carcinoma (HCC) could improve the predictive ability of microvascular invasion (MVI) and pathological grading e ciency [20,21,22].
To our knowledge, no previous studies have built a CT-based radiomics nomogram for HAE. Despite the disadvantage of radiation problem, CT is the most widely used imaging modality for detection in most countries.
This study aims to develop and validate a CT-based radiomics nomogram that would incorporate radiomics signature and clinical factors to evaluate HAE activity.

Patients
This retrospective study was approved by our Ethical Committee, and informed consent was waived for the patients. All procedures involving human participants adhered to the tenets of the Declaration of Helsinki.
The initial 248 patients were reviewed from May 2012 to January 2021. The inclusion criteria were as follows: (1) con rmed HAE by surgery or biopsy; (2) high quality of CT scan were available; (3) PET/ CT were available within 15 days before or after CT scan; (4)complete medical records were at hand, including age, sex, height, weight, body mass index (BMI), and PNM stage (P: parasitic mass in the liver, N: involvement of neighboring organs, M: metastasis) [23].Patients were excluded if any of the inclusion criteria was violated as shown in gure 1. The patients were randomly divided into a training set and test set at the ratio of 8:2. The training set contains 139 HAE patients, of which 99 were active and 40 were inactive. The test set contains 35 HAE patients, of which 25 were active and 10 were inactive.

CT Image Acquisition
All patients underwent a plain scan and contrast-enhanced imaging with a 64-Detector Row CT Scanner (LightSpeed VCT & Discovery 750, GE Medical Systems, USA) with the same scan protocol. CT images were acquired during a single breath-hold. After routine non-enhanced CT, the contrast-enhanced CT scan was initiated after an intravenous administration of 1.5 mL/kg of the iodinated contrast material (Uitravist 370, Bayer HealthCare, Germany) at a rate of 3.0-3.5 mL/s via a high-pressure injector (Tennessee XD2003, Ulrich GmbH & Co. KG, Germany). Three phase-enhanced CT scans were performed, including the arterial phase, portal venous phase, and equilibrium phase. The CT protocol was as follows: volume scan, 120 kVp of tube voltage with automatic tube current modulation, 0.5s of rotation time, 64mm×0.625mm of detector collimation, 5mm slice thickness and interval, 0.984 of pitch, 512mm×512mm of a matrix.

Conventional Radiological Characteristics Analysis and Classi cation
Images were analyzed by two radiologists, both with more than 10 years of experience in diagnosing abdominal diseases. The two radiologists evaluated the images independently over indexes such as signs of microcysts, Graeter classi cation, and calci cation [16]. The calci cation analysis was based on Graeter`s research (1: without calci cation; 2: feathery calci cation; 3: focal calci cation; 4: diffuse calci cation; 5: mainly calci cation, edge calci cation, and central calci cation). One senior radiologist with 15 years of experience in abdomen images was supposed to decide for inconsistent cases.

18F-FDG-PET/CT imaging protocol and image interpretation
All images were obtained from a Discovery VCT PET/CT (GE Healthcare Bio-Sciences, Pittsburgh, PA, USA) with an 18F-FDG tracer produced by Cyclotron (GE Healthcare Bio-Sciences) that had a radiochemical purity of >95%. Patients were intravenously injected with 18F-FDG (7.4 MBq/kg body weight). All patients were treated with standard 18F-FDG-PET/CT acquisition (PET/CT acquisition was performed 1 h after 18F-FDG injection). The delayed 18F-FDG-PET/CT acquisition was performed 3h after 18F-FDG injection if necessary [11]. The image diagnosis (with a de nite "active" or "inactive" label) was made by one experienced radiologists and reviewed by the senior radiologist. The standardized uptake value (SUV) was calculated automatically by semi-quantitative analysis in the workstation. Cases with higher lesion SUV indicates active, lower lesion SUV indicates inactive , compared to that of peripheral liver parenchyma.

Lesion Segmentation and Image preprocessing
The lesion segmentation was performed on the rad cloud platform (version 3.1.0, http://radcloud.cn/, Huiying Medical Technology Co., Ltd, Beijing, China). The volumes of interest (VOIs) were delineated layer by layer along the edge of the lesion in the portal venous phase (the optimal imaging phase for lesion boundary) by one radiologist. All the VOIs were visually con rmed by another senior radiologist. The two radiologists did not know whether the lesions were active when evaluating the imaging. Representative CT images for inactive and active lesions are shown in Figure 2.
To improve the stability of radiomic features, the images were pre-processed by standard deviation normalization [μ-3σ, μ + 3σ], and B-spline interpolation sampling to resample all CT images to1.0×1.0×1.0 mm 3 to unify the slice thickness.

Feature extraction and normalization
The quantitative features of VOIs were calculated at radcloud platform, which was in compliance with de nitions described by the Image Biomarker Standardisation Initiative (IBSI). The IBSI is an international collaboration developed to help standardize radiomic feature calculation, which has made recommendations concerning feature calculation, standardized feature de nition, and nomenclature [24,25]. A total of 1409 radiomics features were extracted, which included rst-order features, shape features, and texture features, which included gray level co-occurrence matrix (GLCM), gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM), neighboring gray-tone difference matrix (NGTDM), and gray level dependence matrix (GLDM). The selected radiomics features of model were in additional le 1 (Table S1).
In feature normalization, Z-score normalization was applied to eliminate the difference in the value scale of the extracted features. The mean value is subtracted from the original feature value and the above results were further divided by the standard deviation.

Features Selection and model construction
To effectively select available features from high dimensional feature sets, the least absolute shrinkage and selection operator (LASSO) logistic regression algorithm was utilized in the training dataset. Radscore for each patient was calculated by using a linear combination of selected features, each weighted by respective LASSO coe cients. The formula used to calculate Rad-score was in additional le 1 (Formula S1).
Supervised machine learning classi ers including K-Nearest Neighbors (KNN), Logical regression (LR), and Multilayer Perceptron (MLP), were applied to develop radiomics models, training in the training dataset and applied in the test dataset. The areas under the receiver operating characteristic (ROC) curves (AUCs) along with 95% con dence intervals (CIs) were used to assess the predictive e ciency of the model, and the cutoff value was selected according to the Youden index to determine the corresponding sensitivity and speci city.

The radiomics and clinical nomogram construction and evaluation
Furthermore, in order to improve the predictive performance of the current Rad-score-based model, the clinical indicators highly correlated with biological activity were introduced into this predictive model. Univariate analysis was rst initiated to select the most highly correlated indicators. A radiomics nomogram was established with both Rad-score and previous selected clinical indicators to predict HAE biological status. Calibration curves and decision curves were delineated to assess the calibration and feasible clinical utility.

Statistical analysis
Software R Version 3.6.3 (https://www.R-project.org/) was used for statistical analysis and models development [25]. For continuous variables, the Kolmogorov-Smirnov test was rst used to evaluate normality. Independent sample t-test was utilized for the data with normal distribution and expressed as mean (standard deviation). Otherwise, the data were analyzed by the Mann-Whitney U test and expressed as median [IQR]. Chi-square test and Fisher's exact probability method were used to compare the categorical variables between groups. DeLong's test was used to compare the differences among models. A two-tailed P < 0.05 was considered to be statistically signi cant.

Result Patients Clinical and Conventional Image Factors
The characteristics of the patients in the training and testing sets were described in Table 1. According to PET/CT indexes, P stage, microcysts sign were signi cantly different (p < 0.05) between the active group and inactive group in the training set and test set; calci cation was signi cantly different (p < 0.05) in training set only; the gender, age, height, weight, BMI, N stage, and M stage were not signi cantly different (p > 0.05) between the active group and inactive group.

Radiomics Models Establishment and Validation
The best performance of LASSO regression was built using a penalty parameter -log(α) = 1.58, as the mean square error was minimized resulted in 30 radiomic features from the 1409 image features (Figure 3). The radiomic features showed good predictive accuracy among three classi ers ( Table 2) according to AUC, sensitivity, and speci city (Figure 4), with MLP achieved the best. In the univariate analysis of the training cohort, the P stage, microcysts sign, and calci cation showed signi cant differences between the active and inactive group. Therefore, a radiomics nomogram model incorporating Rad-score with the clinical indicator was constructed ( gure 5A). Calibration curves for the radiomics nomogram are shown good consistency between predictive outcome and observation in the training and test sets ( gure 5B, C). The results of decision curve analysis (DCA) in the training set are shown in Figure 6. The DCA showed satisfactory performance for the radiomics nomogram model while the proposed radiomics model showed a greater advantage. There were no signi cant differences in the predictive e cacy between the combined models and the single radiomics model.

Graeter Classi cation for Activity Prediction
The results of Graeter classi cation in the active group and inactive group are shown in Table. 3  Discussion HAE is a fatal parasitic disease mainly popular in the temperate countryside and high-altitude mountainous areas. The untreated 10-year mortality was more than 90% [26,27]. PET/CT, the preferred imaging modality, showed the response of in ammatory cells around parasitic lesions, thus indirectly re ecting metabolic activity [10,28]. CT examination was earlier used in HAE evaluation of drug therapy [29], compared with PET/CT, partly because CT equipment is more readily available, especially in these remote pastoral areas. Radiomics based model was not satisfactory compared with clinical indicators based model in predicting the pathological grade or microvascular invasion for HCC [30]. Still, when combined with clinical indicators, the nomogram model becomes better [20]. Therefore, this research was designed to establish a radiomics nomogram with clinical and traditional imaging features to predict HAE biological status.
In this study, three machine learning algorithms were utilized to build prediction models, namely KNN, LR, and MLP, capturing the linear and nonlinear relationships of data, with AUCs ranging from 0.748 to 0.800 in the testing cohort, with MLP achieved the top results. This may be partly due to the fact that the MLP network models the computational units of multiple layers by imitating signal transmission, and the layers of deep neural architecture overcome the limitation of local minimum optimization [31]. The Greater-MLP model which based on CT morphology classi cation, with an AUC of 0.58 and 0.64 in the training and testing cohort. Eventually, a combined nomogram model incorporated the radiomics signature and clinical features while no signi cant differences were detected between the combined model and the radiomics models.
Among clinical markers, the PNM stage indicated the clinical stages derived from the World Health Organization staging system [23]. Our results showed that P stage was related to the AE activity in that with the P stage increased, the nomogram score decreased, implying that a higher the P stage would lead to an inactive lesion. Microcysts sign is one of the most typical image features for HAE. In pathology, the HAE lesions were shown multiple vesicles /microcysts( from 1 mm to 1.1cm in diameter) on gross specimen [2,32]. On CT scan, the small vesicles were represented by small round low density and most clearly displayed in portal vein phase. Previous studies demonstrated that it is related to HAE activity and our results further con rm it [14].
Calci cation is considered as another manifestation of HAE and in most cases, increased calci cation would indicate stable progression [23,33], which is quite similar to Greater's calci cation classi cation [16]. However, Brumpt considered micro calci cations (similar to Greater's feathery calci cation) related to activated status. Clinically, microcalcifcations/feathery calci cations come up with macrocalci cations in most cases, thus indicating that the increase in calci cation maybe a marker of stable progression. Therefore in the current study, Greater classi cation was utilized.
There are still some Limitations to this research. Firstly, this is a retrospective study and the included subjects were heterogeneous, including both untreated and chemotherapy-treated patients, which may cause bias for the results; Secondly, in the delineation of ROI, only portal vein phase images were included in this study and features from CT plain scans and CT enhanced arterial phase images were needed to be veri ed in future; Finally, multicenter research is necessary because there are different characteristics for HAE In different countries and regions [34].
In conclusion, a CT-based radiomics nomogram can evaluate the biological activity of HAE and it is expected to be a more convenient method for follow-up after drug treatment.  Decision curve for models. Y-axis represents the net bene t, which is calculated by gaining true positives and deleting false positives. The X-axis is the probability threshold.