CT-based radiomics models using intralesional and different perilesional signatures in predicting the microvascular density of hepatic alveolar echinococcosis

Abstract

Objectives

To evaluate the performance of CT-based intralesional combined with different perilesional radiomics models in predicting the microvascular density (MVD) of hepatic alveolar echinococcosis (HAE).

Methods

This study retrospectively analyzed preoperative CT data from 303 patients with HAE confirmed by surgical pathology (MVD positive, n = 182; MVD negative, n = 121). The patients were randomly divided into the training cohort (n = 242) and test cohort (n = 61) at a ratio of 8:2. The radiomics features were extracted from CT images on the portal vein phase. Four radiomics models were constructed based on gross lesion volume (GLV), gross combined 10 mm perilesional volume (GPLV_10mm), gross combined 15 mm perilesional volume (GPLV_15mm) and gross combined 20 mm perilesional volume (GPLV_20mm). The best radiomics signature model and clinical features were combined to establish a nomogram. Receiver operating characteristic curve (ROC) and decision curve analysis (DCA) were used to evaluate the predictive performance of models.

Results

Among the four radiomics models, the GPLV_20mm model performed the highest prediction performance with the area under the curves (AUCs) in training cohort and test cohort was 0.876 and 0.802, respectively. The AUC of the clinical model was 0.753 in the training cohort and 0.699 in the test cohort. The AUC of the nomogram model based clinical and GPLV_20mm radiomic signatures was 0.922 in the training cohort and 0.849 in the test cohort. The DCA showed that the nomogram had greater benefits among the three models.

Conclusion

CT-based GPLV_20mm radiomics model can better predict MVD of HAE. The nomogram model showed the best predictive performance.

Introduction

Hepatic alveolar echinococcosis (HAE) is a potentially fatal zoonosis caused by the cestode Echinococcus multilocularis, and predominantly endemic in pastoral regions [1]. Due to its tumor-like growth, HAE has a high morbidity and mortality burden with a 10-year fatality rate of 94% in untreated or inadequately managed patients [2]. The primary treatment of HAE is operative intervention, yet preoperative evaluation of the lesion is essential to determine the feasibility of surgery and to optimize the surgical strategy [3]. The therapeutic goal of HAE is complete removal and killing of hydatid bodies, that is, to render it biologically inactive. Assessment of the bioactivity of the lesions of HAE is crucial to decide the therapeutic strategy and anti-infective drug withdrawal [4]. Studies have shown that the peripheral infiltration zone of HAE lesions has a rich microcirculation similar to that around malignant tumors, which is the active part of HAE lesions and can indirectly reflect the biological activity of worms [5]. Therefore, accurately assessing the MVD of HAE is an important means to evaluate its bioactivity and can provide an important basis for informing clinical treatment decisions and evaluating therapeutic outcomes.

Imaging examination is an important means of preoperative diagnosis and assessment of disease condition and curative effect of HAE. The gold standard for the judgment of the peripheral infiltration zone and bioactivity of HAE is the pathological microvascular density (MVD) in the peripheral area of the lesion, which cannot be morphologically demonstrated by conventional imaging [6]. Several functional imaging modalities, such as fluorodeoxyglucose positron-emission tomography (FDG-PET), contrast-enhanced ultrasound (CEUS), CT perfusion imaging, and quantitative dual-energy CT, have shown promise in evaluating the MVD and bioactivity of HAE lesions [6,7,8]. However, due to the medical conditions and economic constraints, functional imaging methods are difficult to popularize in the pastoral areas where HAE is prevalent. With the rapid development of machine learning-based radiomics, researchers have found new possibilities in quantitative assessment of the MVD and bioactivity of HAE lesions [9,10,11]. Previous studies mostly focused on the extraction of radiomics features within the gross lesion volume. According to the characteristics of chronic outward aggressive growth of HAE, we supposed that the perilesional tissue had a microenvironment more representative of its MVD and bioactivity.

The aim of this study is to construct radiomics models to predict the MVD of HAE based on features extracted from the intralesional and different perilesional regions of HAE, and to find out the best prediction model. Additionally, a visualized nomogram model were constructed with the best radiomics model signature and clinical features to provide a more comprehensive and accurate evaluation for clinical diagnosis and treatment.

Materials and methods

Patients

This retrospective study was approved by the institutional Ethics Committee of the First Affiliated Hospital of Xinjiang Medical University (No. K202312-38) and was conducted in accordance with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. The informed consent was waived as a retrospective study. From January 2012 to December 2023, 467 patients diagnosed with HAE from the First Affiliated Hospital of Xinjiang Medical University were collected. The inclusion criteria were as follows: (1) the MVD of HAE lesions have been pathologically determined; (2) contrast-enhanced abdominal CT images were obtained within 2 weeks before surgery; (3) CT images were met the diagnostic requirements. The exclusion criteria were as follows: (1) clinicopathological record was incomplete; (2) presence of other malignant tumors; (3) preoperative interventional therapy affected the images; (4) the lesion edges were not clearly identifiable for segmentation.

Finally, a total of 303 patients were enrolled. All patients were randomly divided into the training cohort (n = 242) and the test cohort (n = 61) in a ratio of 8:2 (Fig. 1; Table 1). There were 13 features were screened in the preliminary experiment, and the sample size was sufficient according to the 10 events per variable rule.

Fig. 1

Flowchart of the study subjects based on inclusion and exclusion criteria

Table 1 Baseline patient characteristics in training and test cohorts

Pathological diagnosis of MVD and clinical feature selection

Paraffin sections containing HAE lesions and surrounding normal liver tissue were prepared into 4 mm sections. Hbematoxylin-eosin staining and immunohistochemical staining were performed [12]. The MVD count was performed by two trained pathologist with the double-blind method. With reference to Weidner’s method, the “hot spot” of microvascular concentration was first found under low magnification microscope (×40), and then MVD was counted under high magnification microscope (×200) to calculate the mean value as the final MVD [13]. The MVD > 15 was considered positive [14]. Of the 303 patients included in this study, 182 were classified as positive, while 121 were classified as negative.

Clinical features included age, sex, body mass index (BMI), location, number, diameter and calcification type of lesion, some relevant laboratory tests. The calcification type of lesion follows the following criteria: (1) Type A: spotty, minimal calcification with no obvious or discontinuous calcification, or visible spotty calcification that is more limited; (2) Type B: cartographic, moderate calcification, with continuous cartographic calcification in the middle and peripheral zones, with or without central speckled calcification; (3) Type C: large, flaky calcifications, predominantly located in the lateral lesion near the hepatic margin. Laboratory tests mainly include blood routine and liver function related indicators, such as white blood cells (WBC), neutrophils count (NE), lymphocyte count (Lym), aspartate aminotransferase (AST), alanine aminotransferase (ALT), fibrinogen (Fib), platelet count (PLT), creatinine (Crea), total bilirubin (TBIL). Univariate and multivariate logistic regression were used to analyze the clinical features and to select out independent predictors for clinical model construction.

CT image acquisition

All enrolled patients have taken preoperative CT examinations with a multidetector CT system with the following two scanners: Aquilion ONE Genesis Edition (Canon Medical Systems, Japan), GE Discovery HD750 (GE Medical Systems, Milwaukee, WI, USA). During scanning, patients were acquired in the supine position and in inspiratory breath-hold. The scanning direction was craniocaudal direction. Iodized contrast agent (350 mg I/mL, 1.5 mL/kg) was injected through a median cubital vein using a double-ended power syringe (injection speed 3.0 mL/s). Intelligent tracer was adopted in the arterial phase, and scanning starts promptly when CT value of the region of interest (ROI) placed on the abdominal aorta reached 180HU. The portal vein phase and the delayed phase were obtained 50–70 s and 90–120 s after the injection of contrast agent.

Volume of interest (VOI) segmentation and radiomics feature extraction

The workflow of our study was shown in Fig. 2. The VOIs segmentation was performed using 3D Slicer (version 4.13.0). Before segmentation, all the original images were imported into the software and standardized using a window width of 400 HU and a window level of 50 HU. Two radiologists with 5 and 10 years of experience, respectively, who were blind about the patient’s clinical information, depicted the VOIs of gross lesions in portal venous phase as GLV. When there was multiple lesions in the liver, the largest one would be selected. Using the function of dilation as the basis of GLV to expand 10 mm, 15 mm, and20 mm outward, which was recorded as GPLV_10mm, GPLV_15mm, and GPLV_20mm, respectively. The non-liver parenchyma was wiped with the eraser function. The radiomics features were extracted from each VOI using open-source Pyradiomics (version 3.0.1, https://pyradiomics.readthedocs.io/en/latest/index.html) toolbox and processed using Scikit-learn (version 1.0.2, https://scikit-learn.org/stable/index.html) package. Images were resampled at a pixel size of 1 mm×1 mm×1 mm before feature extraction. Radiomics features included first-order features, shape features, gray level co-occurrence matrix features (GLCM), gray level size zone matrix features (GLSZM), gray level run length matrix features (GLRLM), neighbouring gray tone difference matrix features (NGTDM), and gray level dependence matrix features (GLDM). In addition, a variety of filters, such as the wavelet, Laplacian of Gaussian filter (LoG), logarithm, exponential, gradient, square, square root, local binary pattern 2D, local binary pattern 3D, were used to calculate the first-order statistics and texture features of the transformed images.

Fig. 2

Workflow of rasdiomic analysis in this study

Radiomics feature selection and model construction

The radiomics features were standardized using z-score normalization. Interclass and intraclass correlation coefficients (ICCs) were calculated to evaluate the inter-observer and intra-observer agreement for the radiomics features. An ICC > 0.75 of features were considered good and forwarded to the further feature selection. Then the Pearson correlation coefficient > 0.9 was utilized to handle strong correlations between features. To further refine the features, the multivariate least absolute shrinkage and selection operator (LASSO) regression was employed. A 10-fold cross-validation was used to determine the best lambda values to minimize the prediction error. The selected features were subsequently used to develop predictive models using five machine learning algorithms: logistic regression (LR), support vector machine (SVM), random forest (RF), extremely randomized trees (ExtraTrees), and extreme gradient boosting (XGBoost). To ensure the stability and robustness of model performance, a 10-fold cross-validation procedure was repeated over 100 iterations. The classifier with the highest mean area under the curve (AUC) across all iterations was selected as the optimal model.

Nomogram construction

The nomogram based on the logistic regression were constructed with the signature of the best radiomics model and the clinical signature. In addition, the nomogram determined the risk of the MVD by summing the points corresponding to all predictors to facilitate prediction. It provided a direct and visual representation of the assessment for the MVD of HAE.

Statistical analysis

Statistical analyses were performed using R software (version 4.2.2; http://www.R-project.org), and two-side p-values < 0.05 were considered statistically significant. Clinical information was expressed as mean ± standard deviations (SD) or number (frequency). Chi-square test was used for categorical variables. The ICC value of concordance was used to measure the degree of agreement between the two experienced radiologists. The independent predictors of the MVD of HAE were determined by univariable and multivariable logistic regression analysis. The ROC was drawn to evaluate the performance of the model, and difference in AUC was assessed with DeLong test. The DCA curve was used to evaluate the clinical benefit of the models.

Results

Patient characteristics

A total of 303 patients (male/female: 147/156, aged: 38.36 ± 13.11 years) were included (MVD positive, n = 182; MVD negative, n = 121). The flowchart of patients enrolled was shown in Fig. 1. The training and test cohort included 242 and 61 patients, respectively. Baseline patient characteristics in training and test cohorts were shown in Table 1. There were no statistically significant differences in age, sex, BMI, diameter of lesion, WBC, NE, Lym, PLT, Crea, TBIL, ALT and Fib between MVD positive and negative in both training cohort and test cohort (all p > 0.05). Univariate and multivariate logistic regressions showed that calcification type, AST and Fib could be used as independence predictors of MVD positive in HAE(p < 0.05) (Table 2).

Table 2 Logistic regression analysis of clinical factors of patients in the training cohort

Radiomics feature selection

A total of 1874 radiomics features were extracted from the four VOIs, respectively. The consistency of radiomics features calculated by two radiologists was above 0.90. Pearson correlation coefficient greater than 0.90 were considered redundant and removed, leaving 364 features for GLV, 342 for GPLV_10mm, 346 for GPLV_15mm, and 349 for GPLV_20mm. After further dimension reduction by LASSO, the remaining features of the 4 VOIs were 2, 2, 13 and 11, respectively (Fig. 3).

Fig. 3

Selected features weight coefficients. (a) GLV, (b) GPLV_10mm, (c) GPLV_15mm, (d) GPLV_20mm. GLV, gross lesion volume; GPLV_10mm, gross combined 10 mm perilesional volume; GPLV_15mm, gross combined 15 mm perilesional volume; GPLV_20mm, gross combined 20 mm perilesional volume

Radiomics model construction and assessment

As shown in Table 3, the comparison of different algorithms across the four VOI models revealed that the model constructed using the ExtraTrees algorithm achieved a higher AUC in the test cohort compared to models built with other algorithms. It also demonstrated relatively better performance in terms of ACC values. The difference value in AUC of ExtraTrees algorithm between the training cohort and the test cohort was the smallest, indicating that the model was relatively stable. Figure 4 summarized the AUCs of the four VOIs in the training and test cohort. The AUC of GLV, GPLV_10mm, GPLV_15mm, and GPLV_20mm in the training cohort were 0.699 (95%CI, 0.633–0.765), 0.711 (95%CI, 0.645–0.776), 0.806 (95%CI, 0.750–0.861), and 0.876 (95%CI, 0.833–0.918), respectively. The AUC of GLV, GPLV_10mm, GPLV_15mm, and GPLV_20mm in the test cohort were 0.510 (95%CI, 0.356–0.633), 0.691 (95%CI, 0.552–0.830), 0.761 (95%CI, 0.631–0.891), and 0.802 (95%CI, 0.690–0.914), respectively. The GPLV_20mm radiomics model had the best predicted performance among the four VOIs.

Table 3 Comparison of models constructed by different algorithms based on GPLV_20mm optimal features

Fig. 4

AUCs of the four VOIs in the training and test cohort. (a) Training cohort, (b) Test cohort. AUC, area under the curve; VOI, volume of interest

Nomogram performance testing

The Nomogram was constructed by combining the GPLV_20mm radiomics model with clinical signature (Fig. 5). Figure 6a showed the AUCs of three models, in training cohort, the AUCs of clinical model, GPLV_20mm radiomics model, and the nomogram were 0.753 (95%CI, 0.692–0.815), 0.876 (95%CI, 0.833–0.918), 0.922 (95%CI, 0.888–0.956), respectively. As shown in Fig. 6b, the nomogram still performed the highest AUC among the three models in the test cohort, it was 0.849 (95%CI, 0.750–0.947), and the clinical model and GPLV_20mm radiomic model were 0.699 (95%CI, 0.553–0.845), 0.802 (95%CI, 0.690–0.914), respectively. The DeLong test showed that the AUC of monogram was significantly different from the other two models. As shown by decision DCA curves (Fig. 6c-d), nomogram exhibited increased clinical benefit in both training cohort and test cohort.

Fig. 5

Nomogram

Fig. 6

The ROC and DCA curves of the three models for predicting the MVD of HAE. (a-b) ROC in training cohort and test cohort, (c-d) DCA curves in training cohort and test cohort. ROC, receiver operating characteristic Curve; DCA, decision curve analysis

Discussion

This retrospective study extracted perilesional radiomics features from a 0–20 mm range to uncover new insights of the lesion evaluation. The results showed that CT-based radiomics model which constructed by intralesional and 20 mm perilesional region of HAE performed best among the models of four different VOIs, and the nomogram constructed with the radiomic signature and clinical signature shown the highest application value in predicting the MVD of HAE. This is the first study to investigate the performance of radiomics models with different perilesional scales including the 0–20 mm region for predicting the MVD and bioactivity of HAE, and it provided a better understand the infiltration width and safe distance of HAE.

HAE, often referred to as “worm cancer”, exhibits a growth pattern similar to that of liver cancer [15]. Further clinical intervention is based on the peripheral infiltration zone and the bioactivity of HAE [16]. Previous studies have shown that MVD in the infiltration zone of HAE lesions serves as the foundation for their proliferation and development, making it a gold standard for assessing bioactivity [5]. Therefore, preoperative noninvasive assessment of MVD in HAE remains a major clinical challenge. In addition, HAE has a tendency to invade neighboring tissues, which is surrounded by macrophages, lymphocytes, and fibrocytes, and forms granuloma tissues resembling a malignant tumor [17, 18]. The width of its outward invasion is also an important basis for surgical margin. Hillenbrand et al. [19] proposed that long-term disease-free survival in all patients required a safe distance of at least 20 mm provides. From the series of studies of Buttenschoen et al. [20], once the pathologist measured a macroscopic circular minimal safe distance of 17 mm, the microscopic one was only 2 mm. Therefore, they agreed with the view of 20 mm safe distance in previous studies. A study from China has shown that a safe distance of at least 1 mm in combination with medical anthelmintic treatment continuing for two years might offer a good chance of being disease-free long term, but the exact period of medical treatment needed is not defined [19]. Therefore, the width of the marginal infiltration zone is another key factor for clinicians. However, we found that the performance of clinical model in predicting the MVD of HAE was unsatisfactory (training cohort, AUC = 0.753; test cohort, AUC = 0.699).

Radiomics aims to describe and predict the ultrastructural architecture of tumors by analyzing medical images, with the potential to transform radiological tests into virtual biopsies [21]. Machine learning is a scientific discipline focused on how computers learn from data and plays a core technology to achieve artificial intelligence [22]. As radiomics methods and machine learning continue to be applied in the study of various clinical diseases, researchers have begun to explore the feasibility of radiomics in the diagnosis, differentiation, and bioactivity prediction of echinococcosis [9, 10, 23, 24]. Although studies on intralesional radiomics models have achieved certain prediction performance, while are still not accurate and comprehensive. In this study, we first compared the performance of five common machine learning algorithms in predicting the MVD of HAE. The result showed that the ExtraTrees algorithm outperformed the others. The ExtraTrees algorithm is an ensemble method based on the Decision Tree. By dividing completely random features and thresholds, it enhances the diversity of Decision Tree, with reducing over-dependence on specific features in the prediction, which could improve the model robustness and mitigate overfitting. Compared with the machine learning algorithms used in previous studies, it also demonstrates good stability.

The 10 mm range of perilesional radiomics features has been shown to have better predictive efficacy than GLV and clinical models in predicting bioactivity of HAE [25]. However, the maximum range of the edge extracted in previous study was 10 mm. In this study, we extended the range of the VOI to 20 mm, which was recognized as an operational safe distance in clinical practice. Meanwhile, we analyzed radiomics feature models for different VOIs within 10–20 mm margin of the lesion to further explore heterogeneity of the HAE margin zone. The results showed that the GPLV_20mm model exhibited the best diagnostic performance among the four different radiomics models, with the AUCs of the training cohort and test cohort were 0.876, 0.802, respectively. The radiomics features including the surrounding 20 mm of HAE lesion may better predict the MVD of HAE. Additionally, the AUC of GPLV_10mm model were higher than that of the GLV model, which was consistent with the results of previous study. Overall, we can infer that there is still heterogeneity in the tissue surrounding the lesion within 20 mm, which better reflects its peripheral infiltration zone and bioactivity, and also supports the previous clinical finding that the minimum safe distance for HAE radical resection is 20 mm [3].

The study further integrated clinical signature and imaging signature to construct a visual nomogram model. Compared with the clinical model or radiomics model, the nomogram demonstrated the best prediction performance and clinical benefit. The nomogram model, which was more comprehensive and accurate, could provide a reliable individualized basis for the selection of treatment strategies and the evaluation of efficacy of HAE. In recent years, the deep learning technique has also been well applied in various systems of systemic diseases. Two multicenter studies from China have realized automatic identification and classification of HAE using the deep learning-based algorithms with CT plain scan and ultrasound images, respectively [26, 27]. These provided a new inspiration for further studies on the bioactivity and microenvironment of HAE.

Our study still had several limitations. First, it is a single-center study, and this study excluded patients with poor image quality or other treatments were excluded, potentially resulting in selection bias and decreasing the reliability of our study’s conclusions. Second, the correlation between MVD and survival prognosis of HAE has not been investigated. Thirdly, in this study, only radiomics features were extracted for analysis and modeling. In future studies, we will further investigate the heterogeneity and survival prognosis of HAE microenvironment using deep learning and habitat analysis.

Conclusions

CT-based radiomics model which constructed by intralesional and 20 mm perilesional region of HAE was the best radiomics model for predicting its MVD, the nomogram model combined the clinical and radiomics signatures showed the best predictive performance.