2.2. Image Processing and Radiomics Features Extraction

Patients underwent baseline CT scan within 1–3 weeks prior to chemotherapy. During the scan, the patients were requested to suspend their respiration to prevent breathing artifacts. The whole-abdomen CT was performed on patients positioned supine with a slice thickness range of 2–5 mm, and the entire abdomen was scanned (Revolution, GE Healthcare, Milwaukee, WI, USA; Brilliance 64, Phillips, Amsterdam, The Netherlands; SOMATOM, Siemens, Erlangen, Germany). Acquisition and reconstruction parameters were tube current 150–200 mA, tube voltage of 120 kV, pitch 0.8, and matrix size 512 × 512. Section thickness was set at 5 mm and reconstruction section thickness at 1.5 mm. Two experienced diagnostic radiologists performed the segmentation of tumor lesions using ITK-SNAP (version 3.6) software. As the portal lesion was well differentiated from the adjacent tissue, the largest level of the portal lesion was selected for segmentation. Then, segmentation was performed along the border of the tumor, thereby avoiding areas such as the blood vessels and calcifications. If there was a disagreement on segmentation between the two physicians, the two physicians discussed it and reached an agreement.

The CT scan images were initially resampled to a target voxel of 1 mm × 1 mm × 1 mm. Subsequently, the radiomics features were obtained in an automated manner from the manually labeled regions of interest (ROIs) using PyRadiomics software (version 3.0.1), according to the latest recommendations of the image biomarker standardization initiative [11]. Taken together, 1743 radiomic features including 14 shape features, 342 first-order statistics features, and 1387 texture features were extracted from each CT scan with the bin size fixed to 32 [12]. An intraclass correlation coefficient (ICC) analysis was conducted to assess the interobserver reliability of the extracted radiomics features. Radiomics features exhibiting an ICC of 0.75 were considered stable and included in further analysis.

A three-step strategy was used to further select the radiomics features to decrease model complexity and prevent overfitting. Univariate analysis was performed first, and the radiomics features with a significant difference between the C+ and C− groups (Mann–Whitney U test, p < 0.05) were kept. Then, the redundant features were removed via Pearson correlation coefficient analysis. If two radiomics features had a high correlation (|r| > 0.95), the feature with a greater p value as determined using the Mann–Whitney U test was excluded. Finally, the most critical radiomic features were selected using the least absolute shrinkage and selection operator (LASSO). The penalty parameter was established via 10-fold cross-validation using the “minimum mean-square error” method.