CT-Based Radiomics Analysis for Preoperative Diagnosis of Pancreatic Mucinous Cystic Neoplasm and Atypical Serous Cystadenomas

Objectives To investigate the value of CT-based radiomics analysis in preoperatively discriminating pancreatic mucinous cystic neoplasms (MCN) and atypical serous cystadenomas (ASCN). Methods A total of 103 MCN and 113 ASCN patients who underwent surgery were retrospectively enrolled. A total of 764 radiomics features were extracted from preoperative CT images. The optimal features were selected by Mann-Whitney U test and minimum redundancy and maximum relevance method. The radiomics score (Rad-score) was then built using random forest algorithm. Radiological/clinical features were also assessed for each patient. Multivariable logistic regression was used to construct a radiological model. The performance of the Rad-score and the radiological model was evaluated using 10-fold cross-validation for area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy. Results Ten screened optimal features were identified and the Rad-score was then built based on them. The radiological model was built based on four radiological/clinical factors. In the 10-fold cross-validation, the Rad-score was proved to be robust and reliable (average AUC: 0.784, sensitivity: 0.847, specificity: 0.745, PPV: 0.767, NPV: 0.849, accuracy: 0.793). The radiological model performed slightly less well in classification (average AUC: average AUC: 0.734 sensitivity: 0.748, specificity: 0.705, PPV: 0.732, NPV: 0.798, accuracy: 0.728. Conclusions The CT-based radiomics analysis provided promising performance for preoperatively discriminating MCN from ASCN and showed good potential in improving diagnostic power, which may serve as a novel tool for guiding clinical decision-making for these patients.


Inclusion and exclusion criteria
The inclusion criteria were as follows: (i) Surgery was performed at our center, and the pathological diagnosis was made; (ii) Preoperative abdominal CT examination was conducted within two weeks before surgery; (iii) With complete clinicopathological data in electronic medical records. The exclusion criteria were as follows: (i) Preoperative CT was not performed at our center; (ii) Artifact was present in CT images; (iii) With incomplete CT scanning series; (iv) MCN that had progressed to invasive carcinoma. The flow diagram is illustrated in Fig S1.

Image preprocessing, tumor segmentation and feature extraction
Tumor segmentation and image preprocessing were performed via 3D Slicer [1] (version 4.11.0; http://www.slicer.prg). All cases were semi-manually segmented by reader 1 (T.S.X, a junior radiologist with 3-year experience in interpretation of abdominal imaging) on all slices showing the lesion. First, the reader annotated the central area of the lesion and adjacent normal tissue respectively on a transverse, coronal, sagittal section. Then, an approximation algorithm called grow from seeds was used to generate a 3D region of interest (ROI) automatically. Finally, the reader could further optimize ROI manually if necessary.
Before extracting features, each CT scan of each patient was normalized with Z-scores in order to reduce image noise from different CT scanners. Images were then resampled to a voxel size of 1 × 1 × 1 mm 3 . With the Radiomics module in 3D Slicer, 851 radiomics features were extracted based on venous phase CT images. After omitting 87 features with missing values, a total of 764 radiomics features were finally obtained.

Radiomics features information
A total of 764 radiomics features were extracted based on venous-phase CT images, covering six groups (14 shape features, 16 first-order features, 11 gray level dependence matrix features (GLDM), 12 gray level run length matrix features (GLRLM), 12 gray level size zone matrix features (GLSZM), and 699 wavelet-based features). Detailed information about these features is listed in Table S1. Table S1. List of radiomics features classes. GLDM, gray level dependence matrix; GLRLM, gray level run length matrix; GLSZM, gray level size zone matrix. Table S2. Detailed information on radiomics features. GLDM, gray level dependence matrix; GLRLM, gray level run length matrix; GLSZM, gray level size zone matrix.

3.Diagnostic criteria of radiological features.
(i) Tumor size: Tumor size was measured on transverse images. The maximum diameter of the tumor was recorded; (ii) Location: The left margin of the superior mesenteric vein was used to divide pancreatic head/neck and body/tail; (iii) Lesion contour: Lesion contour was split into two groups named round/ovoid and lobulated. If the lesion contour could not be described as the borders of the same circle, it was defined as a lobulated contour, or it was defined as a round/ovoid contour; (iv) Wall thickness: If the wall thickness of a cystic lesion was larger than 2 mm, it was considered thick. If the wall was imperceptible or wall thickness smaller than 2 mm, it was considered thin; (v) Wall enhancement: Wall enhancement was described as line or rim-like enhancement of lesion wall in arterial or venous-phase CT images. The enhanced wall was visible for at least 50% of the lesion circumference; (vi) Calcification: Calcification presented on the wall or septa was noted as positive; (vii) Mural nodules: A nodule-like structure on the inner side of the wall or on the septa was noted as mural nodules positive; (viii) Dilation of the Wirsung duct: The main pancreatic duct was considered dilation if its diameter was larger than 2 mm. Examples of radiological features are illustrated in Fig S2  Fig S2. Examples of radiological features. An asterisk was used to annotate the target lesion.

4.Reproducibility analysis for radiomics feature extraction.
A total of 764 radiomics features were extracted for each patient. The first step of the feature selection was reproducibility analysis. Thirty patients were randomly chosen to evaluate the inter/intra-observer intraclass correlation coefficient (ICC) of radiomics features. Reader 1 (T.S.X, a junior radiologist with 3-year experience in interpretation of abdominal imaging) repeated segmentation flow twice with time intervals exceeding one week; reader 2 (X.Y.W, a junior radiation oncologist with 4-year experience in irradiation volume segmentation of abdominal tumor) independently performed segmentation flow; the intraclass correlation coefficient (ICC) was calculated to evaluate feature reproducibility. All readers were blinded to the pathological diagnosis during segmentation.
Among all 764 radiomics features, there were 491 and 492 features with interobservation ICC and intra-observation ICC higher than 0.90, respectively (Fig S3.). Finally, 472 features with both inter-and intra-observation ICC higher than 0.90 were selected for the next step. Features with good agreement are annotated by a blue bar, while a poor agreement is annotated by a red bar.

R packages in this study
The R packages used in this study were as follows. Random forest algorithm was completed using the 'randomForest' package. The 10-fold cross validation was performed using the 'caret' package. ROC was printed using 'pROC' package. The logistic regression model was constrcuted using the 'rms' package. The calculation of ICC was done by using the 'irr' package.