Prediction of high Ki-67 proliferation index of gastrointestinal stromal tumors based on CT at non-contrast-enhanced and different contrast-enhanced phases

The study was approved by the review board of Renji Hospital and the need for written informed consent was waived. This is a retrospective single-center study conducted at the Department of Radiology in Renji Hospital, using consecutive patients diagnosed with pathologically confirmed GIST from July 2012 to June 2022. The inclusion criteria were as follows: (1) CT images were available before pathological analysis, including at least one non-contrast-enhanced and three contrast-enhanced phases in a single CT examination, and (2) tumors were pathologically confirmed as GIST by fine needle aspiration or surgical specimens. The exclusion criteria consisted of (1) comorbidity of primary tumors other than GIST, (2) lacking Ki-67 PI in immunohistochemistry data, and (3) tumor that were either partially included in CT images or had suboptimal CT imaging quality. The most recent CT scan was selected if multiple contrast-enhanced CT images were available for a single subject. This study adheres to the Image Biomarker Standardisation Initiative (IBSI) guidelines for the standardization of image acquisition and analysis.

All subjects underwent a non-contrast-enhanced abdominal CT scan and a three-phase contrast-enhanced CT scan using one of the specified CT systems, with the scanning and reconstruction parameters used in daily clinical practice. The detailed CT protocols are presented in Table S1. The iodine contrast agent was administered at a dose of 1 mL/kg of the patient’s weight, and the flow rate was 2.5–3.5 mL/s. The three phases in the contrast-enhanced CT images included arterial phase (25–30 s after injection), venous phase (70–80 s after injection), and delayed phase (180–240 s after injection).

All CT images were imported to 3D Slicer (version 5.0.3, https://​www.​slicer.​org/​) for delineation. The delineated areas included the tumor lesion as much as possible without including the surrounding normal tissues or other tissues to generate a volume of interest (VOI). Interpolation (“Fill between slices”) and smoothing tool were utilized in order to make the VOI boundary smooth and accurately match the real lesion. CT image segmentation was conducted by one radiologist (Z.X.) with 3 years of radiological experience, delineated either at the delayed phase or the phase with the clearest tumor boundary in the contrast-enhanced CT. Subsequently, a simple transformation, including translation, rotation, resizing along the axial axis, or a combination of these, was manually applied to align the delineated VOI with the lesion in the non-contrast-enhanced phase and other contrast-enhanced CT phases. In rare cases, where the tumor shape was distinct in different phases, manual segmentation was performed in all four CT phases without transformation. The image segmentation process was repeated 2 months later by the same radiologist (Z.X.) on a randomly selected subset of 20 patients, and another radiologist (W.Z.) with 5 years of radiological experience performed image segmentation once on the same subset. Additional details on the conversion to mask was described in the Supplementary Materials. Two-way mixed effects, consistency, single rater/measurement intra- and inter-observer correlation coefficient were then applied. Intra- and inter-observer correlation coefficient estimates were calculated using Python software, version 3.10.8 (Python Software Foundation) based on a single-rater, absolute-agreement, two-way random-effects model [16]. The features with an intra- and inter-observer correlation coefficient greater than 0.75 were then selected for analysis.

All CT images were obtained using CT scanners from the following vendors: GE (n = 218), Philips (n = 19), Siemens (n = 96), Toshiba (n = 29), and UIH (n = 21). CT images obtained from GE scanners were used as the training dataset. Two validation datasets were created to address the imbalanced sample size of the vendors: one from Siemens (validation dataset 1), and the other from the remaining vendors (n = 69, validation dataset 2). These validation datasets were reserved and remained untouched until the model was fine-tuned.

A total of 107 radiomics features were extracted from the VOI of each phase of CT images without applying additional filters. After filtering based on intra- and inter-observer correlation coefficient, 79 features were selected for use in model fine-tuning. These features, which included first-order statistics, shape, grey-level co-occurrence matrix (GLCM), grey-level dependence matrix (GLDM), grey-level run-length matrix (GLRLM), grey-level size-zone matrix (GLSZM), and neighboring grey tone difference matrix (NGTDM), were extracted using PyRadiomics, version 3.0.1 in Python software, version 3.10.8 (Python Software Foundation), compliant to IBSI definitions. Details about the radiomics extraction workflow are included in the Supplementary Materials. To avoid over-fitting, a limited number of radiomics features with higher predictive performance to the target variable were selected. Through exploratory data analysis, we found a high inter-feature correlation among most radiomics features. Thus, we first ranked all radiomics features based on their Spearman correlation coefficient with the value of Ki-67 PI, and included the top 4 features with inter-feature correlation coefficient maintained less than 0.9. This process was repeated for non-contrast-enhanced CT and all phases of contrast-enhanced CT.

The random forest models were trained for each phase of CT images to predict tumors with a high Ki-67 PI (Ki-67 > 10%). The optimization of the random forest models was performed using the Bayesian optimization algorithm in a given parameter search space. The hyperparameters with the best average area under receiver operating characteristic curve (AUC) over three-fold cross-validation were used as the optimal parameters to fit the entire training set. This process was repeated for both the non-contrast-enhanced phase and all three contrast-enhanced phases. The binary prediction was directly made by the random forest models, and the predictive performance of the models was then evaluated on both validation datasets.

Descriptive statistics were presented as frequencies (n) and percentages (%) for categorical variables and mean ± standard deviation for continuous variables. Comparisons of categorical variables between different groups were performed using Fisher’s exact test or Pearson’s χ² test with Yate’s correction for continuity. Continuous variables were compared using either Students’ t test or one-way analysis of variance (ANOVA), as appropriate. The DeLong test was used to calculate the variance of AUC and to compare the performance of different models. Randomized permutation test with 1000 permutations was used to determine the significance of the model predictions. A two-sided p value of < 0.05 was considered statistically significant. All statistical analyses were performed with Python software, version 3.10.8 (Python Software Foundation). The source code for our paper was publicly available at https://​github.​com/​heyrict/​gist-eus-open. Data in our paper was available upon reasonable request to the corresponding author.

Our study has several limitations. Firstly, our study was a single-center retrospective study with a limited sample size. To minimize selection bias, we utilized a multi-vendor approach in validating our models. In addition, there was a gap between AUC values in the two validation datasets, which may result from the limited sample size, or possibly a scanner dependency problem. Although we are confident in the generalizability of our models as the DeLong test is not significant and all of them achieved an AUC > 0.7, a larger, prospective, multi-center trial is still needed to confirm these results. Secondly, the sample size of CT images obtained from different vendors was unevenly distributed; thus, we combined images from several minor vendors into one validation dataset to mitigate potential biases. In addition, the Ki-67 PI of several samples were acquired from fine needle aspiration specimens, which could lead to potential bias considering the heterogeneity in Ki-67 PI values. Nevertheless, a previous study has demonstrated good correlation between FNA-based Ki-67 PI and mitotic count in surgical specimens, and we have determined that fine needle aspiration can be used for Ki-67 PI assessment in GISTs [27]. Lastly, our use of three-dimensional VOI for tumor segmentation is time-consuming and subject to debate regarding inter-reader variability. However, previous research has shown that three-dimensional texture analysis offers a more comprehensive representation of tumor heterogeneity in abdominal tumors [28], and we achieved an acceptable intra- and inter-observer correlation in our radiomics features (79 over 107 features with both intraclass correlation coefficient > 0.75).