Our study constructed a multi-modal CT and MRI radiomics prediction model of CT and MRI features, which proved to have the potential ability in predicting 2-year OS of ESCC patients. It showed that our radiomics model contributed more than the clinical model to the stratification of low-risk and high-risk patients in terms of overall survival. To our knowledge, this is the first study to predict the OS of ESCC patients using integrated CT and MRI radiomics features.
Compared to MRI-based radiomics, most results in EC radiomics come from positron emission tomography-computed tomography (PET-CT) and CT images.[7, 22–25] CT is the most basic imaging modality in the staging of EC. The CT images can be expediently obtained and re-used to extract deeper information. Therefore, CT-based radiomics has made great progress in predicting the prognosis of patients with EC. Xie et al. extracted sub-region based radiomics features from EC patients receiving (chemo)radiotherapy and confirmed that the sub-region based radiomics model had AUC values of 0.821 in the training group and 0.805 in the validation group in predicting 2-year OS, higher than the AUC values of 0.733 and 0.654 in our study.[20] This may be due to the small number of patients in their experiment, only 87 in the training group and 46 in the validation group. Tang et al. further investigated that the combination of radiomics and clinical features had a better performance than either of them in predicting early recurrence of locally advanced ESCC (AUC values of combined group vs radiomics group vs clinical group in validation groups: 0.809 vs 0.646 vs 0.658).[24] A multicenter study further developed and validated the hybrid radiomics nomogram of radiomics signatures, deep-learning signature and clinical factors in predicting local recurrence of ESCC patients after definitive (chemo)radiotherapy (C-index in training, internal validation and external validation set: 0.82 vs 0.78 vs 0.76).[26] However, our study found that the predictive performance of clinical factors was lower than that of radiomics. Combining clinical factors with radiomics would significantly reduce the accuracy of the prediction model. This may be due to the fact that the radiomics features had much better prediction performance than the clinical features, so adding clinical factors did not improve the accuracy of the model. Therefore, we did not include clinical factors in the integrated model.
Unlike the flourishing CT radiomics studies, MRI-based radiomics researches were relatively rare and in their infancy. Hirata et al. explored the relationship between apparent diffusion coefficient (ADC) related features and prognosis and demonstrated that histogram analysis of ADC could predict recurrence-free survival and disease-specific survival in ESCC patients.[13] Chu et al. constructed a combined model of MRI-based radiomics and clinical features and showed high accuracy in predicting OS (C-index in training and validation groups: 0.730 and 0.712) and DFS (C-index: 0.714 and 0.729).[27] However, in our study, the clinical factors are slightly insufficient in predicting prognosis, but the combination of the two imaging radiomics has better predictive efficacy. MRI has been shown to be very useful in determining the T stage, particularly when the tumor is not clearly demarcated from the trachea and great vessels in EC.[28, 29] The incorporated MRI and CT radiomics can provide additional information about tumor biological characteristics and heterogeneities, which are associated with the prognosis of patients with rectal cancer.[30, 31] Li et al. demonstrated that an MRI-based radiomics model was more effective than CT in predicting therapeutic response after neoadjuvant chemotherapy for locally advanced rectal cancer. The combination of MRI and CT radiomics achieved the highest AUC value of 0.925 in the training group and 0.93 in the validation group.[15] However, no similar study on hybrid MRI and CT radiomics has been conducted in EC. MRI is rarely used in the diagnosis of EC due to respiratory movement and heartbeat affect the sharpness of MRI imaging. Additionally, the coils used to scan the cervical and chest esophagus are different, making it possible to scan the entire esophagus once.[32] However, advancements in MRI technology now allow for long-range and whole-body MRI scans which are possible and clinically available.[33, 34] Moreover, techniques such as the ultrasound-driven 4D MRI method and sensor systems have been developed for respiratory motion imaging and respiratory gating in thorax and abdomen scans.[35, 36] Additionally, MRI has certain advantages in staging T3 and T4 patients and determining resectability for surgeons in ESCC, making it an increasingly important tool in the diagnosis and treatment of EC.
This study found that the hybrid model had the best predictive ability compared to the MRI or CT models, respectively. Additionally, the MRI model performed slightly better than the CT model in predicting the 2-year OS of ESCC patients after dCRT, achieving an AUC of 0.715 in the validation group, which is better than some mono-modal image studies.[7, 25] Moreover, the radiomics models were far more accurate than the screened clinical factors in terms of prediction accuracy. It is hypothesized that the pre-treatment imaging captured more detailed and individual features of the patient group, thus better reflecting the tumor heterogeneity and providing the predictive value. Furthermore, the hybrid radiomics model was used to significantly stratify high-risk and low-risk patients, which has a profound guiding significance for the follow-up treatment of ESCC patients.
Recent studies have shown that combining raidomics with biomarkers such as HER2 and CD44 may lead to promising results in predicting prognosis.[37] Xie’s study suggested a significant correlation between copy number alterations (CNA) and radiomics features.[20] The combination of genomics and radiomics may have greater predictive potential in clinical practice and may reveal potential biological pathways of radiomics. In recent years, it has been confirmed that certain features in the image of pathological tissue can be used to predict survival in non-small cell lung cancer.[38] In rectal cancer, a combination of histopathological and radiomics features can predict tumor response better than three single-modality prediction models in terms of AUC values (0.812, 95%CI 0.717–0.907 VS 0.630, 95%CI 0.507–0.754; 0.716, 95%CI 0.580–0.852; 0.733, 95%CI 0.620–0.845).[39] In the future, the combination of multi-disciplinary omics may provide more prognostic information for survival prediction and model construction.
Our study has some limitations in our study. Firstly, it was a single-center clinical trial and lacked external validation. Secondly, the small number of patients included in the study and the limited image data available for analysis may lead to inaccuracies in the results. Finally, to ensure high repeatability of images and positions, we acquired images of MRI positioning on the same day after CECT positioning. Therefore, we were unable to obtain contrast-enhanced MRI images, and as a result, we did not analyze the data from the enhanced MRI phase. Moreover, we did not include the ADC calculated from the DWI phase in our analysis. However, further analysis of this information in the future may improve the accuracy of the model’s prediction.
In conclusion, the results indicated that the multi-modal radiomics of MRI and CT improved the 2-year OS prediction ability in ESCC patients compared to either modality alone and outperformed the clinical features and can better stratify patients with different risks.