A contrast-enhanced CT-based whole-spleen radiomics signature for early prediction of oxaliplatin-related thrombocytopenia in patients with gastrointestinal malignancies: a retrospective study

Patient population

This retrospective observational study was approved by our institutional review board (approval number: TJ-IRB20201221) and informed patient consent was waived. Patients with pathologically or cytologically diagnosed gastrointestinal malignancy who were admitted to the Cancer Center of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology from March 2017 to December 2020 were reviewed. Patients were included if they were over 18 years of age and chemotherapy naive, received a minimum of two cycles and a maximum of twelve cycles of oxaliplatin-based chemotherapy (either first-line or adjuvant, including modified FOLFOX6 (mFOLFOX6), CAPOX and SOX regimens) and underwent follow-up contrast-enhanced abdominal CT within three months after the start of chemotherapy. Patients with baseline platelet count <100*10⁹/L or received concurrent radiotherapy were excluded. Oxaliplatin-based chemotherapy regimens were listed in Table S1. The final study population consisted of 119 patients. Patients were randomly divided into the training (n = 85) and validation (n = 34) cohort with a ratio of 7:3.

Patients were followed up regularly and laboratory tests, including complete blood count were monitored before and during each cycle of chemotherapy. Thrombocytopenia during chemotherapy was recorded and was graded as per National Cancer Institute’s Common Terminology Criteria for Adverse Events (CTCAE; version 5.0) based on platelet count as follows: grade1, ≥75*10⁹/L and <100*10⁹/L; grade 2, ≥50*10⁹/L and <75*10⁹/L; grade 3, ≥25*10⁹/L and <50*10⁹/L; grade 4, <25*10⁹/L (National Cancer Institute, 2017). According to CTCAE, no intervention is indicated for grade 1 thrombocytopenia. Therefore, in this study we focused on grade 2 or higher thrombocytopenia, since more intensive medical observation and intervention are required in these cases. Patients were divided into two groups, the thrombocytopenia group (TCP Group) and the non-thrombocytopenia group (Non-TCP Group), based on whether they developed grade 2 or higher thrombocytopenia during chemotherapy. Follow-up contrast-enhanced abdominal CT were performed every 1.5–2 months after the start of chemotherapy. Demographic, clinical, laboratory and imaging data of participants were recorded. The workflow of this study was shown in Fig. 1.

CT image acquisition

All contrast-enhanced abdominal CT scans were acquired on a 64-slice CT system (Discovery C750 HD; GE Healthcare). The following parameters were used: tube voltage, 100–120 kV; tube current, 200–350 mA; slice thickness, 5 mm; matrix, 512*512. Non-ionic contrast medium Iopromide (Ultravist, Bayer Healthcare, Wayne, NJ, iodine concentration, 370 mg/mL) was injected intravenously at a rate of 3 ml/s followed by a 25 ml saline flush. Bolus tracking technique was used to automatically trigger arterial phase, portal venous phase and delayed phase acquisition 5–8 s, 20–25 s and 180–240 s after the attenuation of abdominal aorta reached 150 HU, respectively. The raw data were reconstructed into 1.25 mm slice thickness and exported in DICOM format for analysis.

Spleen segmentation, volume measurement and radiomics feature extraction

3D Slicer (version 4.11.20210226, http://www.slicer.org), an open-source software was used to segment spleen, measure volume and extract radiomics features from the first follow-up contrast-enhanced abdominal CT images. A radiologist with 5 years of experience in abdominal imaging semi-automatically segmented the whole spleen on portal venous phase images using the Nvidia AI-Assisted Annotation (Nvidia AIAA) extension package implemented in 3D Slicer. After several boundary points had been set near the edge of spleen on axial, sagittal and coronal views, 3D Slicer software segmented the whole spleen within a minute, the radiologist then interactively corrected the segmentation if necessary and large blood vessels and lesions were carefully excluded. An example of the semi-automatic whole spleen segmentation was showed in Fig. S1. All segmentations were further double-checked by another radiologist with 10 years of experience in abdominal imaging. The volume of spleen was calculated after semi-automatic segmentation. Finally, a total of 851 radiomics features, consisted of first-order features (n = 18), shape features (n = 14), second- and higher-order features (n = 75) and wavelet features (n = 744) were extracted using the SlicerRadiomics extension package implemented in 3D Slicer. To determine the intra- and inter-reader reproducibility of radiomics features, 20 randomly-selected cases were segmented by the same radiologist again after a period of 1 month and by another radiologist with 10 years of experience in abdominal imaging.

Radiomics feature selection and radiomics signature development

Before feature selection, Z-score normalization was performed to rescale all radiomics features in a standard normal distribution in the training cohort. The mean and standard deviation of each radiomics feature in the training cohort were used for the normalization in the validation cohort. Radiomics features with intraclass correlation coefficients > 0.75 in intra- and inter-reader reproducibility tests were considered reproducible and included in feature selection. The independent two-sample t-test and the Mann–Whitney U test were then conducted to identify statistically significant radiomics features between the TCP Group and the Non-TCP group in the training cohort. The least absolute shrinkage and selection operator (LASSO) logistic regression algorithm was applied to further select features strongly associated with the development of grade 2 or higher thrombocytopenia in the training cohort. LASSO is a simple dimensionality reduction method with smaller mean squared error and deals with multicollinearity problems and overall variable selection, therefore is widely used in a variety of model fitting (Lee et al., 2014). The glmnet package in R software (version 3.6.1; R Core Team, 2019) was used for LASSO regression with penalty parameter tuning conducted by 10-fold cross-validation. A Radiomics signature, radiomics score (Rad-score) was constructed by a linear combination of the selected features weighted by their respective LASSO coefficients. The diagnostic performance of radiomics signature was evaluated through receiver operating characteristic (ROC) curve analysis in the training and validation cohorts. The robustness and stability of radiomics signature was evaluated by repeated 10-fold cross validation (200 times) in the whole cohort.

Clinical model and clinical-radiomics model development and performance evaluation

Clinical factors including age, gender, body mass index (BMI), primary tumor site, liver metastases, cumulative dose of oxaliplatin, number of oxaliplatin-based chemotherapy cycles, oxaliplatin-based chemotherapy regimen, concurrent bevacizumab treatment and spleen volume were analyzed using univariable logistic regression in the training cohort. Clinical factors with p < 0.05 were included into multivariable logistic regression to build models predicting the development of grade 2 or higher thrombocytopenia in the training cohort. A clinical model that included only clinical factors, as well as a clinical-radiomics model that incorporated clinical factors and radiomics signature were built. Multicollinearity was assessed using variance inflation factor (VIF) before multivariable logistic regression. A nomogram of the clinical-radiomics model was constructed to provide a visual, intuitive and individualized prediction of the development of grade 2 or higher thrombocytopenia during oxaliplatin-based chemotherapy.

Model performances including discrimination and calibration of both clinical and clinical-radiomics models were evaluated to determine the incremental value of radiomics signature over conventional clinical factors. The area under the ROC curve was calculated to assess the discrimination of models. The bootstrap method was performed to compare the area under the curve (AUC) of both models. The Hosmer–Lemeshow goodness-of-fit test and the Akaike information criterion (AIC) were used to assess the calibration of the models (Steyerberg et al., 2010). A non-significant p value of the Hosmer–Lemeshow test indicates good calibration. The AIC is a statistic evaluating the goodness-of-fit as well as the simplicity of a prediction model. A lower AIC value indicates a better-fit model. Empirically, if the AIC difference between models (ΔAIC) > 2, the model with the lower AIC value is considered substantially better fitted (Hu et al., 2021). Integrated discrimination improvement (IDI) index was calculated to evaluate the incremental prognostic accuracy after adding radiomics signature over conventional clinical factors for risk prediction of thrombocytopenia (Pencina et al., 2008).

Statistical analysis

Statistical analysis was performed with R software (version 3.6.1; R Core Team, 2019). The Chi-square test or Fisher’s exact test were used to compare categorical variables while the independent two-sample t-test and the Mann–Whitney U test were used to compare continuous variables between the training and validation cohorts. Two-tailed p < 0.05 was considered statistically significant.

Patient characteristics

A total of 119 patients (70 men; median age, 54 years; interquartile range, 50–64 years) with gastrointestinal malignancy, including 74 patients with colorectal cancer and 45 patients with gastric cancer were included into this study. Before chemotherapy 24 patients (20.2%) presented with synchronous liver metastases. The mean BMI of all patients was 21.50 ± 2.65 kg/m². All patients received at least two cycles of oxaliplatin-based chemotherapy (range, 2–12 cycles; median number of cycles, 6). The median cumulative dose of oxaliplatin was 780 mg/m² (range, 255–1,040 mg/m²). The majority of patients (75 out of 119, 63.0%) was treated with CAPOX regimen. About 10.1% of patients received bevacizumab in combination with CAPOX (7.5mg/kg every 3 weeks, n = 4) or mFOLFOX6 (5 mg/kg every 2 weeks, n = 8) regimen. The baseline characteristics of patients in the training and validation cohorts are summarized in Table 1. No significant difference was found between these two cohorts.

Follow-up

Thrombocytopenia was observed in 68 out of the 119 patients (57.1%) during oxaliplatin-based chemotherapy. Slightly more than half of the patients who developed thrombocytopenia (37 out of 68, 54.4%) were mild cases (grade 1) and no medical intervention was required. Grade 2 or higher thrombocytopenia occurred in 31 patients (26.1%), with a median time interval of 101 days (interquartile range, 70–141 days) from the start of chemotherapy to the onset of grade 2 or higher thrombocytopenia. The incidence of grade 3 (five out of 119, 4.2%) and grade 4 (two out of 119, 1.7%) thrombocytopenia was low and no bleeding episode was reported.

All patients underwent contrast-enhanced abdominal CT after oxaliplatin-based chemotherapy. The median time interval between the start of chemotherapy and the first follow-up CT scan was 49 days (interquartile range, 43–65 days), ranging from 34 to 90 days. The median spleen volume at follow-up was 215.94 cm³ (interquartile range, 170.65–273.89 cm³).

Radiomics signature construction and validation

After feature selection, LASSO algorithm identified five radiomics features with non-zero coefficients, including wavelet-LHL_firstorder_Energy, wavelet-HHL_GLCM_Idn, wavelet-HHL_GLSZM_HighGrayLevelZoneEmphasis, wavelet-LLL_firstorder_Kurtosis and wavelet-LLL_NGTDM_Complexity, which were strongly associated with the development of grade 2 or higher thrombocytopenia in the training cohort (Fig. 2). Rad-score was calculated for each patient using a linear combination of the 5 identified features multiplied by their respective coefficients. Calculation formula is presented in Supplementary Material. A significant difference in Rad-score was found between the TCP Group and the Non-TCP Group in the training cohort (−0.697 ± 0.357 vs −1.227 ± 0.325, p < 0.001) and further verified in the validation cohort (−0.900 ± 0.456 vs −1.288 ± 0.290, p =0.006). Patients in the TCP Group had significantly higher Rad-scores. The ROC curves indicated good performance of Rad-score in predicting grade 2 or higher thrombocytopenia in both cohorts (Fig. 3). The average AUC in the whole cohort from repeated 10-fold cross validation was 0.839 ± 0.146. Using the optimal cut-off value determined by maximizing the Youden’s index in the training cohort, Rad-score demonstrated favorable sensitivity, specificity and accuracy in the training cohort (81.8%, 84.1%, 83.5%, respectively) as well as the validation cohort (77.8%, 80.0%, 79.4%, respectively).

Clinical model and clinical-radiomics model development and performance evaluation

Univariable logistic regression revealed that the cumulative dose of oxaliplatin, baseline BMI, liver metastases and spleen volume at follow-up were significantly associated with the development of grade 2 or higher thrombocytopenia. Patients who received higher cumulative dose of oxaliplatin, with liver metastases, lower baseline BMI and greater spleen volume at follow-up were significantly more likely to develop grade 2 or higher thrombocytopenia during chemotherapy. By contrast, age, gender, primary tumor site, chemotherapy regimen, number of chemotherapy cycles and bevacizumab treatment failed to show significant impacts on the risk of developing grade 2 or higher thrombocytopenia (Table 2).

A clinical predictive model and a clinical-radiomics predictive model were constructed using the statistically significant clinical factors in the univariable analysis alone and together with radiomics signature respectively (Table 2). Baseline BMI, cumulative dose of oxaliplatin and spleen volume were the 3 factors included in the clinical model. Rad-score and cumulative dose of oxaliplatin were the only 2 factors in the clinical-radiomics model. The clinical-radiomics model was presented as a nomogram in Fig. 4. As shown in Fig. 5 and Table 3, compared to the clinical model, the clinical-radiomics model manifested higher sensitivity, specificity and accuracy in discriminating patients who developed grade 2 or higher thrombocytopenia from those who did not in both training and validation cohorts. In the training cohort, the AUC of the clinical-radiomics model and clinical model was 0.913 (95% CI [0.851–0.974]) and 0.828 (95% CI [0.720–0.935]), respectively. In the validation cohort, the AUC of the clinical-radiomics model and clinical model was 0.867 (95% CI [0.727–1.000]) and 0.751 (95% CI [0.564–0.938]), respectively. In the whole cohort, the average AUC from repeated 10-fold cross validation was higher for the clinical-radiomics model compared with the clinical model (0.882 ± 0.106 vs 0.784 ± 0.154) though not statistically significant (bootstrap method, p = 0.057). However, IDI index showed that integrating radiomics signature to conventional clinical factors significantly improved the predictive accuracy by 17.0% (95% CI [4.9%–29.1%], p = 0.006) in the whole cohort. The Hosmer-Lemeshow goodness-of-fit test yielded non-significant p values (clinical model, p = 0.537 (training cohort) and 0.369 (validation cohort), and clinical-radiomics model, p = 0.893 (training cohort) and 0.257 (validation cohort)), suggesting favorable calibration of both models. However, the AIC value of the clinical-radiomics model was substantially lower than that of the clinical model (62.316 vs 81.789, ΔAIC = 19.473), suggesting a significantly better calibration of the clinical-radiomics model. The calibration curves of the two models in the training and validation cohorts were showed in Fig. 6. Figure 7 showed the clinical-radiomics nomograms of two cases for predicting the risk of developing grade 2 or higher thrombocytopenia during oxaliplatin-based chemotherapy.