Contrast-enhanced CT radiomics features to preoperatively identify differences between tumor and proximal tumor-adjacent and tumor-distant tissues of resectable esophageal squamous cell carcinoma

Abstract

Background

Esophagectomy is the main treatment for esophageal squamous cell carcinoma (ESCC), and patients with histopathologically negative margins still have a relatively higher recurrence rate. Contrast-enhanced CT (CECT) radiomics might noninvasively obtain potential information about the internal heterogeneity of ESCC and its adjacent tissues. This study aimed to develop CECT radiomics models to preoperatively identify the differences between tumor and proximal tumor-adjacent and tumor-distant tissues in ESCC to potentially reduce tumor recurrence.

Methods

A total of 529 consecutive patients with ESCC from Centers A (n = 447) and B (n = 82) undergoing preoperative CECT were retrospectively enrolled in this study. Radiomics features of the tumor, proximal tumor-adjacent (PTA) and proximal tumor-distant (PTD) tissues were individually extracted by delineating the corresponding region of interest (ROI) on CECT and applying the 3D-Slicer radiomics module. Patients with pairwise tissues (ESCC vs. PTA, ESCC vs. PTD, and PTA vs. PTD) from Center A were randomly assigned to the training cohort (TC, n = 313) and internal validation cohort (IVC, n = 134). Univariate analysis and the least absolute shrinkage and selection operator were used to select the core radiomics features, and logistic regression was performed to develop radiomics models to differentiate individual pairwise tissues in TC, validated in IVC and the external validation cohort (EVC) from Center B. Diagnostic performance was assessed using area under the receiver operating characteristics curve (AUC) and accuracy.

Results

With the chosen 20, 19 and 5 core radiomics features in TC, 3 individual radiomics models were developed, which exhibited excellent ability to differentiate the tumor from PTA tissue (AUC: 0.965; accuracy: 0.965), the tumor from PTD tissue (AUC: 0.991; accuracy: 0.958), and PTA from PTD tissue (AUC: 0.870; accuracy: 0.848), respectively. In IVC and EVC, the models also showed good performance in differentiating the tumor from PTA tissue (AUCs: 0.956 and 0.962; accuracy: 0.956 and 0.937), the tumor from PTD tissue (AUCs: 0.990 and 0.974; accuracy: 0.952 and 0.970), and PTA from PTD tissue (AUCs: 0.806 and 0.786; accuracy: 0.760 and 0.786), respectively.

Conclusion

CECT radiomics models could differentiate the tumor from PTA tissue, the tumor from PTD tissue, and PTA from PTD tissue in ESCC.

Background

Esophageal cancer is one of the most common malignant tumors threatening human health, and its mortality and morbidity rank 6th and 7th in the world, respectively [1]. Esophageal squamous cell carcinoma (ESCC) and adenocarcinoma are the main histological types, and ESCC accounts for approximately 90% of esophageal cancer cases worldwide [2, 3]. The overall 5-year survival rate of ESCC ranges from 15%—25% [4]. Currently, esophagectomy remains the mainstay treatment for ESCC [5], but in some patients with histopathologically negative margins, there is still a relatively higher recurrence rate. Previous studies revealed that field cancerization (FC) is a crucial factor in tumor recurrence [6, 7]. The notion of FC is that molecular changes in histologically normal adjacent tissues are similar to the tumor itself, which could lead to new invasive carcinomas in the resection margins or in “macroscopically normal” tissues adjacent to the tumor [7,8,9,10]. In addition, some previous studies reported FC approximately 1 cm away from the tumor margins (tumor-adjacent tissues) and near the ideal resection margins (tumor-distant tissues) in esophageal cancer patients, and found differences in gene expression and molecular mechanisms between tumor-adjacent and tumor-distant tissues [11, 12]. If the tissue showing precancerous lesions is not completely removed by surgery, it may cause local recurrence or develop into a second primary tumor [13]. In addition, transthoracic esophagotomy plus proximal partial gastrectomy were the main surgical methods for ESCC, and esophagogastric anastomosis was used to restore the continuity of the digestive tract. Therefore, it is of vital importance to preoperatively identify differences between the tumor, proximal tumor-adjacent (PTA) and proximal tumor-distant (PTD) tissues to formulate optimal surgical strategies to reduce tumor recurrence.

Contrast-enhanced computed tomography (CECT) is the preferred imaging method for the preoperative diagnosis of ESCC and treatment decision making. However, CT is mainly used to evaluate the external morphological characteristics of tumors, rather than assessing intratumor heterogeneity. Radiomics can quantitatively analyze a large number of features extracted from traditional medical images, such as CT, and noninvasively provide potential information about the biological characteristics and internal heterogeneity of tumors [14, 15]. In recent years, radiomics has been widely used to noninvasively evaluate the preoperative staging, chemoradiotherapy response and postoperative recurrence of ESCC [16,17,18]. To the best of our knowledge, no study has assessed the differences between the tumor, PTA and PTD tissues of ESCC using CECT radiomics. Therefore, the purpose of this study was to develop and validate radiomics models based on CECT to preoperatively differentiate between the tumor, PTA and PTD tissues of ESCC.

Methods

Patients

The institutional ethics committee of our institution approved this retrospective study, and written informed consent was waived.

From July 2020 to June 2022, we collected CT data and medical records of 553 consecutive patients with ESCC from Centers A and B. Patients were recruited into this study based on the following inclusion criteria: (a) patients underwent preoperative thoracic CECT scans within 2 weeks before surgery, and did not receive preoperative neoadjuvant chemotherapy and/or radiotherapy before undergoing CT examination; (b) ESCC was confirmed by preoperative endoscopy and pathological biopsy, and the surgical margin was postoperatively confirmed to be negative; (c) the patients had no history of other organ malignancies; and (d) the upper margin of the tumor was at least 5 cm away from the cricopharyngeal muscle. The exclusion criteria were as follows: (a) the clinicopathological information was incomplete (n = 4); (b) the quality of CT images was unsatisfactory (n = 6); (c) the patients had surgical contraindications (n = 6); or (d) the primary tumor of ESCC could not be identified on CECT (n = 8). Finally, 24 of 553 patients were excluded, and a total of 529 patients were involved in our study. Among the 529 patients, 447 patients from Center A were randomly assigned to the training cohort (TC, n = 313) and the internal validation cohort (IVC, n = 134) at a 7:3 ratio based on the published report [19], and 82 patients from Center B were assigned to the external validation cohort (EVC). The patient flowchart is displayed in Fig. 1.

Fig. 1

Flowchart for selecting the study population. Notes: CECT, contrast-enhanced computed tomography; ESCC, esophageal squamous cell carcinoma

All enrolled patients with ESCC received radical esophagectomy and regional lymph node dissection. The surgical margins in all resected specimens were not affected by the tumor (R0). Baseline clinical data of all patients, including age, sex, anatomical site of the primary tumor, differentiation degree, cT stage and cN stage, were obtained from the medical records (Table 1). Additionally, the differentiation degree, cT stage and cN stage were determined based on postoperative histopathological biopsy and the eighth edition American Joint Committee on Cancer guidelines [20]. In addition, patients with advanced ESCC determined by postoperative pathology received adjuvant treatments such as radiotherapy or chemotherapy.

Table 1 The clinical characteristics of patients from the Centers A and B

Follow-up

We followed up all enrolled patients every 3 to 6 months for one year or more after surgery through barium swallow, thoracoabdominal CT imaging or endoscopic biopsy, focusing on suspicious recurrence or the corresponding clinical symptoms related to postoperative recurrence, such as dysphagia. In addition, patients with confirmed recurrence after surgery underwent adjuvant treatments (such as radiotherapy or chemotherapy) to prolong their survival period. In each participant, the termination event of the follow-up was postoperative death during the follow-up period after surgery. The related survival data of all patients are also presented in Table 1.

Image acquisition

All patients enrolled in our study underwent thoracic CECT scans with two 64 multidetector scanners (LightSpeed VCT, GE Medical systems, USA; and SOMATOM Definition AS + , Siemens Healthineers, Erlangen, Germany). Before CT data acquisition, 100 to 200 mL of water was needed to be drunk as oral esophageal negative contrast material. CT examinations were executed during one breath hold at full suspended inspiration for 10–15 s in the supine position. After a conventional unenhanced CT scan, CECT data acquisition was started 25 to 30 s after the initiation of 1.5 mL/kg nonionic iodine contrast agent (containing 300 mL/kg iodine) at a rate of 3.0 mL/s for a total of 70 to 100 mL via a 20-G needle inserted into an antecubital vein with a high-pressure injector and flushed with 20 mL saline. The parameters of the two CT scanners were as follows: tube voltage of 120 kV, tube current of 200 mA, detector collimation of 64 × 0.6 mm, reconstruction slice thickness of 1 mm, matrix of 512 × 512 mm for each scanner, rotation time of 0.5 s and 0.3 s, and pitch of 0.9 and 0.8 for the corresponding scanner. The anatomic coverage of the CT scan was from the chest entrance to the middle level of the left kidney. All the CECT data used for radiomics feature extraction were retrieved from the picture archiving and communication system.

Image segmentation and radiomics feature extraction

To accurately extract the radiomics features of ESCC, PTA and PTD tissues, the definitions of the previous three tissues on CECT imaging were as follows. For ESCC, the esophageal wall thickness exceeding 5 mm was considered abnormal due to the primary tumor on axial imaging [21]. Clinically, the ideal proximal resection margin for ESCC is ≥ 5 cm [11, 22], so this study only investigated the characteristics of the proximal resection margin of ESCC. At a distance of 1 cm away from the tumor proximal margin, the molecular variation related to cancer is obvious, and this variation can also extend to the resection margin [11]. Therefore, we defined PTA and PTD as the tissues approximately 1 cm and 5 cm away from the proximal margin of the tumor, respectively.

The thoracic CECT images of ESCC with 1 mm thickness were imported into 3D-Slicer (https://www.slicer.org), a free and open-source software, which was used for delineating regions of interest (ROIs) of the tumor (Fig. 2A), PTA and PTD tissues, and extracting radiomics features of the corresponding tissues. The locations of the PTA and PTD tissues were shown on the reconstructed sagittal CECT (Fig. 2B), and the axial ROIs of the PTA and PTD tissues were obtained at the corresponding distance (Fig. 2C). Two radiologists (readers 1 and 2, with 4 and 25 years of experience in digestive radiology, respectively) who were blinded to the patients’ pathological outcomes separately delineated ROIs of ESCC slice by slice and the ROIs of the PTA and PTD tissues (the area of each ROI more than half of the esophageal wall) on the magnified axial CECT scans, avoiding fat, air, bone and blood vessels. When two readers disagreed with each other, they achieved consensus after discussion and consultation. Finally, 1223 individual radiomics features of ESCC, PTA and PTD tissues were automatically extracted by applying the 3D-Slicer radiomics module, and radiomics features included first-order, gray-level co-occurrence matrix (GLCM), gray-level dependence matrix (GLDM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighboring gray tone difference matrix (NGTDM), and shape.

Fig. 2

In a 64-year-old male patient with esophageal squamous cell carcinoma (ESCC), the region of interest (ROI) was delineated slice by slice on axial contrast-enhanced computed tomography (CECT) images (A). The white arrow refers to the primary tumor of ESCC; the blue, red and green solid lines represent the position lines of the proximal tumor-distant (PTD) tissue, proximal tumor-adjacent (PTA) tissue and the proximal margin of ESCC on the reconstructed sagittal CECT image, respectively; and the yellow arrow along the longitudinal axis of the esophagus indicates the ruler measuring line for defining the boundary lines of the PTD and PTA tissues (B). The ROI covering more than half of the esophageal wall for PTA tissue is manually drawn on an axial magnified CECT image (C). The ROI delineation of PTD tissue is performed similarly to that of PTA tissue

Intra- and interobserver agreements

To assess the repeatability of radiomics feature extraction, we randomly chose 60 consecutive samples from Center A to test the intra- and interobserver consistency by the previous 2 radiologists. The intraclass correlation coefficient (ICC) was used to evaluate the intra- and interobserver agreements. An ICC score greater than 0.75 indicated satisfactory agreement. To assess the intraobserver reproducibility, reader 1 sketched the ROIs of the tumor, PTA and PTD tissues twice within one week following the same delineating steps based on a published report [19]. In addition, reader 2 independently delineated the ROIs of the tumor, PTA and PTD tissues to assess the interobserver agreement by comparing the extracted radiomics features from the first ROI depiction by reader 1.

Dimensionality reduction and radiomics feature selection

To avoid the curse of dimensionality and reduce the bias from radiomics features when modeling [19], we adopted the following three steps to select the significant features of the pairwise tissues (ESCC vs. PTA, ESCC vs. PTD, and PTA vs. PTD).

First, all the previous 1223 radiomics features underwent z-score normalization [23]: ${x}_{norm}=\frac{x-\mu }{\sigma }$, where x is the original feature value, µ is the mean value of this feature, and σ is the standard deviation.

Second, all features were examined by Student’s t test or the Mann‒Whitney U test to select potentially significant features from TC. Radiomics features that did not satisfy either of the aforementioned tests (P > 0.05) were excluded.

Third, the least absolute shrinkage and selection operator (LASSO) was used to identify the core radiomics features in TC to determine the differences between the previous pairwise tissues by performing variable selection and regularization of high-dimensional data, thus enhancing the accuracy and interpretability of the core radiomics features [19, 24]. The 1-standard error of the minimum criteria (the 1-SE criteria, a simpler model) was applied to adjust the regularization parameter (λ) for feature selection using 10-fold cross-validation.

Construction and validation of the radiomics model

The optimal selected radiomics features from TC were used to build radiomics models based on logistic regression, a classical machine learning method. The radiomics models were used to evaluate the diagnostic performance of core radiomics features in TC for differentiating ESCC from PTA or PTD tissue and PTA from PTD tissue. The corresponding radiomics models of IVC and EVC were also obtained through the aforementioned logical regression method to validate the performance of the above radiomics models from TC. The area under the receiver operating characteristic curve (AUC), accuracy, F-1 score, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated by the confusion matrix to evaluate the radiomics models based on the previous individual pairwise tissues of TC, IVC and EVC.

Statistical analysis

All statistical analyses of radiomics data were conducted using R statistical software (version 4.2.1, https://www.r-project.org/). The “psych” package was used to evaluate the intra- and interobserver agreements of all radiomics features extracted from the corresponding ROIs of previous pairwise tissues of ESCC. LASSO regression based on multivariate logistic regression was performed using the “glmnet” and “pROC” packages, and the “pROC” package was used to plot the receiver operating characteristic (ROC) curves of TC, IVC and EVC. A P value less than 0.05 suggested a significant difference.

Radiomics quality score

The readers 1 and 2 assessed the methodologic quality of our radiomics study in consensus through the radiomics quality score (RQS) proposed by Lambin in 2017 [25]. The RQS comprises 6 domains with 16 components. Domain 1, the assessment of the quality and replicability of image and segmentation; domain 2, the reporting of the feature reduction; domain 3, model performance and validation; domain 4, biological validation and potential clinical utility; and domains 5 and 6, demonstration of high-level evidence and open science, respectively [26]. Most items are designated to 0, 1 or 2 points. In order to highlight the importance of some dimensions, a higher point is assigned. For example, a prospective validation study is assigned 7 points, and a study validated in three or more datasets is assigned 5 points, while a study without validation is assigned -5 points. The ideal score of the RQS is 36 points, corresponding to a percentage of 100%.

Results

Intra- and interobserver agreements of feature extraction

For intra- and interobserver agreements of CECT radiomics feature extraction from the 60 random consecutive samples, there were 1122, 1137 and 1126 extracted features of the pairwise tissues (tumor vs. PTA, tumor vs. PTD, and PTA vs. PTD) with ICC values greater than 0.75, whereas there were 101, 86 and 97 unsatisfactory features with ICC values less than or equal to 0.75, respectively (Fig. 3). After this assessment, the 1122, 1137 and 1126 radiomics features with all intra- and interobserver ICC values of more than 0.75 were selected for further analyses. Reader 1 performed the segmentation and feature extraction of the remaining participants.

Fig. 3

Feature stability evaluation with intra- and interobserver agreements based on the intraclass correlation coefficient (ICC). All features show good intraobserver agreement (A, B and C) and interobserver agreement (D, E and F) with ICCs > 0.75 (above the red cutoff line) in pairwise comparisons, including esophageal squamous cell carcinoma vs. proximal tumor-adjacent (PTA) tissue, tumor vs. proximal tumor-distant (PTD) tissue, and PTA vs. PTD tissue, respectively

Dimensionality reduction and radiomics feature selection

Among the previous 1122, 1137 and 1126 features in the pairwise tissues, including tumor vs. PTA, tumor vs. PTD, and PTA vs. PTD in all ESCC patients, the Student t test or Mann‒Whitney U test showed that 1013, 1050 and 823 features were statistically significant, respectively (all P values < 0.05). These features with statistical significance by the previous tests were separately used for the subsequent LASSO regression. For the pairwise tissues, including tumor vs. PTA, tumor vs. PTD, and PTA vs. PTD, 20, 19 and 5 features were screened (Tables 2, 3 and 4) to perform the corresponding differentiation, respectively, and 10-fold cross-validation was applied to select the best tuning regularization parameter (λ) under the 1-SE criteria (Fig. 4).

Table 2 The 20 features to differentiate tumor from proximal tumor-adjacent tissue of esophageal squamous cell carcinoma

Table 3 The 19 features to differentiate tumor from proximal tumor-distant tissue of esophageal squamous cell carcinoma

Table 4 The 5 features to differentiate PTA from PTD of esophageal squamous cell carcinoma

Fig. 4

Radiomics feature selection using least absolute shrinkage and selection operator (LASSO) regression in pairwise tissues. The area under the receiver operating characteristic curve (AUC) was plotted by tuning the optimal parameter (λ) selection via 10-fold cross-validation and minimum criteria in the LASSO model. The left and right dotted lines denote the minimum criteria and the 1-standard error criterion (1-SE) in the pairwise tissues, including esophageal squamous cell carcinoma vs. proximal tumor-adjacent (PTA) tissue, the tumor vs. proximal tumor-distant (PTD) tissue, and the PTA vs. PTD tissue (A, B and C, respectively). The LASSO coefficient profiles the 1013, 1050 and 823 radiomics features for the differentiation of previous pairwise tissues, and as a result, 20 19 and 5 nonzero coefficients have been produced (D, E and F, respectively)

Construction and validation of the radiomics models

Through logistic regression, the above 20, 19 and 5 selected core radiomics features were applied to build individual radiomics models for identifying the differences between the tumor and PTA tissue, the tumor and PTD tissue, and the PTA and PTD tissues in TC, respectively. Then, the 3 models for the individual pairwise tissues were validated by the IVC and EVC. The most appropriate individual models selected by AUC, accuracy, F-1 score, sensitivity, specificity, PPV, and NPV are illustrated in Table 5. The ROC curves (Fig. 5) visually revealed that the individual radiomics models had excellent ability to differentiate the tumor from PTA tissue and the tumor from PTD tissue in TC, and effective performance for the previous differential diagnoses in the IVC and EVC. As shown in Fig. 6, the logistic radiomics model showed good performance for differentiating the PTA from PTD tissue in TC and was also helpful for distinguishing the PTA from PTD tissue in the IVC and EVC.

Table 5 Performance of the radiomics models to differentiate between tumor, PTA and PTD of ESCC

Fig. 5

The receiver operating characteristic (ROC) curves of the radiomics models for differentiating between esophageal squamous cell carcinoma and proximal tumor-adjacent tissue and between the tumor and proximal tumor-distant tissue in the training cohort (A and B, respectively), internal validation cohort (C and D, respectively), and external validation cohort (E and F, respectively)

Fig. 6

The receiver operating characteristic (ROC) curves of the radiomics model for differentiating between proximal tumor-adjacent tissue and tumor-distant tissue in the training cohort (A), internal validation cohort (B) and external validation cohort (C)

The RQS of our radiomics study

According to the RQS evaluation criteria and reporting guidelines, the RQS score of our study was 16, accounting for 44.4% of the total points. The scores for each item are listed in Table 6.

Table 6 The assessment of the radiomics quality score

Discussion

In the current study, we initially developed and validated individual CECT radiomics models for the preoperative differentiation of tumor from PTA tissue, tumor from PTD tissue, and the PTA from PTD tissue in ESCC.

As shown in this study, 1122 and 1137 candidate radiomics features with ICCs > 0.75 in the pairwise tissues (including ESCC vs. PTA and ESCC vs. PTD) were reduced to 20 and 19 core features to develop the individual radiomics models to effectively differentiate the tumor from PTA tissue and the tumor from PTD tissue of ESCC, and the previous core features were composed of 13 and 14 texture features, 3 and 2 first-order features, and 4 and 3 shape features, respectively. Some selected features, such as skewness in the first-order feature class, appear to be related to the differentiation of the tumor from PTA or PTD tissue; however, it is challenging to reliably link a single radiomics feature with the complex pathological state of different tissues [27]. Therefore, constructing and validating multifeature models is a more feasible and reliable approach to potentially differentiate the tumor from PTA tissue and tumor from PTD tissue. Our CECT radiomics models showed excellent diagnostic performance in TC, IVC and EVC, with all AUCs of more than 0.95, indicating that our radiomics models could play a pivotal role in differentiating the tumor from PTA tissue and the tumor from PTD tissue.

Moreover, we also selected 5 core features from the 1126 features with ICC > 0.75 in the pairwise tissue (PTA vs. PTD) to develop the CT radiomics model to differentiate between the PTA and PTD tissues of ESCC. The 5 core features include 1 texture feature, 1 first-order feature and 3 shape features. The texture feature and the first-order feature most likely reflect the differences in microarchitecture and internal heterogeneity between the PTA and PTD tissues. Shape features exhibit the external contour information of the PTA and PTD tissues. Our study revealed that the radiomics model developed with the previous 5 core features could help differentiate the PTA from PTD tissue in TC (AUC: 0.870), as validated by IVC (AUC: 0.806) and EVC (AUC: 0.786). For the first time, our study developed a CT radiomics model for discriminating the PTA from PTD tissue of ESCC, and we can speculate that differentiation in radiomics features between the PTA and PTD tissues might contribute to preoperative evaluation of the extent of surgical resection to some degree.

In our research, the following methods are adopted to ensure robustness. First, to ensure the accuracy of radiomics feature extraction, we drew the ROI along the tumor edge based on a reference criterion that the thickness of the abnormal esophageal wall was more than 5 mm [28, 29], which has been proven reasonable. For PTA and PTD tissues, we depicted the area of each ROI in at least more than half of the esophageal wall to obtain sufficient CECT image information. Second, intra- and interobserver agreements, univariate analysis and LASSO were used for feature selection, thus guaranteeing the independence and accuracy of each feature extraction in the final models. We used 10-fold cross-validation and stepwise regression to avoid overfitting of the data and guarantee the robustness of the models [24]. Third, our study added EVC to validate the performance of the radiomics models in TC, which ensured the reliability and stability of the models for differentiating the tumor from PTA tissue, the tumor from PTD tissue, and the PTA from PTD tissue. Fourth, RQS was used to assess the methodologic quality of our radiomics research, and our study obtained a score of 16, greater than or equal to the scores of several relevant articles in the published meta-analysis review [30], indicating our current radiomics models could help guide the design of our future radiomics study to improve the reliability and repeatability of radiomics models for clinical application.

There are still several limitations in our study. First, our study is a retrospective study. A prospective study will be needed in the future to further verify our results. Second, we depicted the ROIs of PTA and PTD tissues at only two levels (1 cm and 5 cm away from the proximal margin of the tumor) and did not investigate the differences in radiomics features of the surrounding-tumor tissues at different sections away from the proximal margin of the tumor. We will conduct relevant research in the future. Third, we did not consider the proteomic and metabolomic characteristics at present. P53 protein accumulation and metabolic profile generation in the tumor and proximal histologically normal tissue have been proven to be evidence for field cancerization in patients with esophageal cancer [12, 31]. We will combine radiomics with proteomics or metabolomics for our future studies. Last but not least, an existing study [32] has shown that cancer biology might evolve during neoadjuvant treatment, resulting not only in a chemoradiotherapy-resistant residue but also possibly a more aggressive cancer. In addition, we did not know whether neoadjuvant therapy would impact the radiomics features of the tumor or the PTA and PTD tissues of resectable ESCC in patients with advanced cancer. Therefore, our study only enrolled patients who did not undergo neoadjuvant therapy to ensure the accuracy and reliability of the extracted radiomics features of the tumor and the PTA and PTD tissues of resectable ESCC.

Conclusions

Our study innovatively developed and validated individual CECT radiomics models using core radiomics features in pairwise tissues (including tumor vs. PTA, tumor vs. PTD, and PTA vs. PTD), and the radiomics models were proven to be valuable in differentiating the tumor from PTA tissue, the tumor from PTD tissue, and the PTA from PTD tissue. We hope that our radiomics models could be helpful for formulating surgical decision-making to reduce postoperative local recurrence.