Predicting the efficacy of neoadjuvant therapy in ESCC patients is crucial for the development of personalized treatments. Although previous studies have attempted to predict the efficacy of treatments for certain diseases using biomarkers, there are currently no definitive biomarkers for immunotherapy in ESCC [25–27]. Research has shown that utilizing tumor imaging data can provide more direct insights [28–31]. Considering the accessibility and health economics of examinations, as a recommended method for preoperative tumor staging in esophageal cancer treatment guidelines[32], enhanced CT scans are regarded as a valuable tool. They can identify tumor changes, facilitate preoperative assessments, and predict treatment efficacy. This approach offers a macroscopic and direct method to assess tumor characteristics in patients with ESCC. Previous radiomics studies [28–31] have only analyzed the intra-tumoral characteristics of ESCC patients, However, features should not be limited to the sole internal region of the tumor. Recent research [33–35] suggests that surrounding areas can provide complementary information on tumor heterogeneity. Therefore, we propose a radiomics model based on enhanced CT images, which combines intra-tumoral and peri-tumoral radiomics features to predict the efficacy assessment of ESCC patients after receiving NICT.
In this study, we comprehensively analyzed 1967 radiomics features extracted from intra-tumoral and peri-tumoral regions obtained from enhanced CT images. Subsequently, we screened out image features that can predict the efficacy of neoadjuvant therapy in esophageal cancer patients. Ultimately, we identified five most significant features from both intra-tumoral and peri-tumoral regions. Three features belonged to gray-level size zone matrix (GLSZM) features. GLSZM quantifies regions of gray levels in images, defined as the number of connected voxels sharing the same gray level intensity. One feature belonged to first-order features, describing voxel intensity distribution within the image region defined by a mask. Another feature belonged to gray-level co-occurrence matrix (GLCM), a method for describing texture features by studying the correlation of different gray values at specific angles and distances within an image. The classification results of these features suggest a potential correlation between patient GR state and tumor heterogeneity, as features based on first-order statistics, GLCM, GLSZM, neighborhood gray-tone difference matrix (NGTDM), gray-level run length matrix (GLRLM), and gray-level dependence matrix (GLDM) are generally considered to reflect tumor heterogeneity at both global and local scales[36].
Wu et al. [37] previously, extracted 10 intra-tumoral features from CT images of 154 patients for analysis. Results showed that some features could differentiate early (Stages I-II) and advanced (Stages III-IV) ESCC, with respective training cohort areas under the receiver operating characteristic curve (AUC) of 0.795 and 0.694, and internal validation cohort AUCs of 0.762 and 0.624. In contrast, our results showed that combining intra-tumoral and peri-tumoral tissue in the predictive model yielded AUC values of 0.809 and 0.800 on the training and internal validation cohorts, respectively, this suggests that the peri-tumoral region may provide complementary useful information, thereby potentially enhancing the predictive ability of the model. This finding is partially consistent with recent research results[38–41].
Dong et al. [42] and Liu et al.[43]previously demonstrated the importance of clinical T-stage as a predictive indicator. Through statistical analysis of clinical features, we also found that T-stage is an important predictor of GR status, negatively correlated with GR status in ESCC patients. The AUC for T-stage was 0.809, 0.800, and 0.716 for the training cohort, internal validation cohort, and external validation cohort respectively. The Rad-Score for predicting response to NICT achieved an AUC of 0.838 for the training cohort, 0.831 for the internal validation cohort and 0.831 for the external validation cohort. By combining T-stage with the established Rad-Score model, we established a clinical imaging radiomics nomogram, which exhibited the best predictive performance. the AUC values for the training cohort, internal validation cohort, and external validation cohort were 0.867, 0.871, and 0.818 respectively. Outperforming both standalone radiomics and clinical T-stage models, indicating improved ability to detect GR in patients. DCA indicated that our nomogram model could provide more benefits for ESCC patients in predicting GR. The visual approach of the nomogram compared to machine learning algorithms may assist doctors in diagnostic decision-making. Predictive models can non-invasively and accurately predict the efficacy of NICT, providing valuable choices and suggestions for personalized treatment of ESCC patients, thus alleviating the burden and suffering of patients during treatment and holding significance for prognosis and treatment prediction.
This study has limitations. Firstly, it is a retrospective study with a relatively small sample size. In the future, using more clinical data to validate our model could enhance its robustness. Secondly, the manual delineation of ROI introduces variability despite the involvement of radiologists, leading to poor repeatability of imaging data and time-consuming operations. Incorporating deep learning-based lesion automatic or semi-automatic segmentation holds promise for improving this aspect in our future work. Thirdly, there is a lack of follow-up data regarding patients' survival time, which hinders the analysis of subsequent features and parameters.