4. Discussion
Radiation resistance is an independent poor prognostic factor, and adverse reactions affect tumor response and long-term survival for LACC. This study integrated clinicopathological parameters and treatment planning CT-based radiomics for predicting CR, 5-year OS, and hematologic toxicities. The clinical and radiomic models outperformed the single models (validation balanced accuracies: 80.79% vs 70.34% and 75.24%) in predicting CR, indicating that a hybrid approach may have greater power for CR. For predicting OS, the radiomic model showed superior performance (validation balanced accuracy: 85.75%). Although only using the tumor volume to build SVM models can predict short-term CR with good performance (validation balanced accuracy: 72.87%), it showed low predictive power for long-term OS (validation balanced accuracy: 50%). To predict hematologic toxicities, combining radiomic features with clinicopathological parameters can achieve validation balanced accuracies of 69.93% and 70.96% for severe leucopenia and hypohemoglobin, respectively. Treatment planning CT radiomics of bone marrow may be potential biomarkers for the prediction of treatment outcomes and hematologic toxicities in LACC.
Previous studies have demonstrated that tumor volume, pelvic lymph node status, and concurrent chemotherapy are independent prognostic factors of LACC [35]. But Lucia et al. reported that radiomic models have higher prognostic power than clinicopathological parameters, such as FIGO stage, tumor volume, and nodal stage [36]. Fang et al. constructed a multihabitat MRI radiomic model to predict tumor response with an AUC of 0.8 [37]. Our models integrated the CT radiomic features with clinicopathological parameters to predict CR achieved an AUC of 0.87, which outperformed the clinical (AUC 0.75) model, CT radiomic model (AUC 0.85), and MRI model (AUC 0.80). For predicting 5-year OS, our radiomic model reached an AUC of 0.82, the clinical and radiomic model only achieved an AUC of 0.81. The predictive power of the joint model is reduced possibly because different treatment regimens were performed to 257 patients. After the models were re-trained using 159 patients treated equally, the clinical and radiomic model achieved better performance than the radiomic model (Table S3). Previously, Ho et al. and Lucia utilized MRI radiomics to predict disease-free survival for cervical cancer patients, but the absence of long-term survival led the both studies to be limited [36, 38]. We predicted both short-term CR and long-term survival, which help radiation therapists distinguish radio-resistance candidates in the early stage and adjusted treatment regimen in time such as the addition of radio-sensitizers and/or more intensive follow-up.
Bone marrow cells are easily damaged by low-dose radiation, which may be associated with hematologic toxicities. Previous studies have found the correlation between hematologic toxicities and dose-volume parameters of pelvic bone marrow based on PET-CT in rectal cancer and gynecological oncology [39, 40]. Utilizing CT radiomics of the pelvis and sacral vertebrae to predict hematologic toxicities has not yet been reported. We combined clinicopathological parameters and CT radiomics to predict hypohemoglobin (grade > 0) and severe leucopenia (grade ≥ 3) with AUCs of 0.74 and 0.64. The decreased hemoglobin levels are associated with the prognosis of radiotherapy and hypoxia-induced radio-resistance [41, 42]. Severe leucopenia increases the risk of infection and radiotherapy often needs to be suspended when it happens, which negatively affects the therapeutic efficacy. The accurate and timely prediction of hematologic toxicities may help to avoid serious complications. Future studies will take dosimetric factors of pelvis and sacral vertebrae into account to further improve the predictive power of our models for hematologic toxicities.
There are some limitations in our study. (i) Our results were not externally validated. External and multicentric data are needed to validate our results. (ii) The radiomic predictors for hematologic toxicities were extracted from CT images, which cannot fully present bone function alterations and bone-related diseases. Other imaging techniques, such as bone mineral densitometry, PET, and multiparametric MRI, may provide more predictive information for analyzing radiation-induced hematologic toxicities. (iii) We only studied planning CT and contrast-enhanced MRI T1-weighted images for predicting treatment outcomes without considering other CT/MRI sequences and modalities that may reveal underlying information for prediction. Further research is needed to investigate multimodality radiomic models and refine the optimal combination of clinical and imaging radiomic features.
In conclusion, noninvasive models based on clinicopathological parameters and treatment planning CT radiomics had predictive power for CR, 5-year OS, and hematologic toxicities. Before clinical application, external validation with a larger cohort is needed to refine the models for predicting the risk of treatment failure and hematologic toxicities.