Prediction of distant metastasis in esophageal cancer using a radiomics–clinical model

Based on clinical factors and radiomics, this study developed and validated a nomogram prediction model for distant metastasis of esophageal cancer with high discrimination and robustness.

The main clinical predictors of distant metastasis were age, pathological differentiation, and N stage. The clinical prediction model performed well, with an AUC of 0.731, which was consistent with our previous research [23]. The radiomics–clinical model was more accurate after the addition of radiomics, with an AUC = 0.827, demonstrating that radiomics can supplement clinical risk factors in the prediction of distant metastasis of EC.

Radiomics played an important role in predicting tumor metastasis, because medical images can show molecular phenotypes of tumors from a macro-perspective [24, 25]. The goal of radiomics is to convert images into data that can be mined, extracted, and analyzed [26]. The radiomic features of primary lesions can help predict lymph node metastasis of esophageal cancer. Qu et al. screened texture features from MRI images, concentrating on length, shape, gray-level co-occurrence matrix (GLCM), and gray-level run length (GLRL). The radiomic signature created by these features can accurately determine whether esophageal cancer patients have lymph node metastasis with an AUC = 0.821, (95% CI 0.7042–0.9376) [27]. In a retrospective study of 230 patients with esophageal cancer, Zhang et al. discovered that CT-based radiomics can be used to predict lymph node metastasis, which is more accurate than simply using lymph node size as the judgment standard [28]. Radiomics of esophageal cancer can be used as a biomarker to predict radiotherapy and chemotherapy efficacy [29,30,31]. The application of radiomics in the treatment of prostate, lung, and breast cancer has also been studied [12,13,14,15]. According to Tunali et al., radiomic features capture biological and pathological information, which has been shown to provide rapid and noninvasive biomarkers for lung cancer risk prediction, diagnosis, prognosis, and tumor biology [32].

In this study, the Lasso-logistic regression algorithm was used to screen high-dimensional radiomic features, and 16 features associated with distant metastasis and without multicollinearity were chosen. The majority of them were texture features or texture features after wavelet transformation, such as the gray-level co-occurrence matrix (GLCM), gray-level size zone matrix (GLSZM), gray-level dependence matrix (GLDM), and neighborhood gray tone difference matrix (NGTDM). Wu et al. filtered 10 radiomic features from 6140 to distinguish early esophageal cancer from advanced esophageal cancer. These features were primarily found in GLCM, GLRLM, GLSZM, and NGTDM [33], which were consistent with our findings. KNN, RF, SVM, and LR algorithms were used to build models based on these selected features to evaluate the maximum efficiency of radiomics in predicting distant metastasis of esophageal cancer and select the best algorithm for fitting the features. The results demonstrated that the prediction ability of radiomics models constructed by machine learning algorithms other than RF was comparable to that of clinical factors for the prediction of esophageal cancer distant metastasis and can be used as a marker to predict EC distant metastasis on its own.

A multi-feature-based radiomics signature can provide more information than a single parameter [21]. The combination of radiomic signature and clinicopathological factors through machine learning can optimize the performance of prediction models [34, 35]. In this study, the efficiency of the radiomics–clinical model (AUC = 0.827) was significantly higher than that of the clinical model (AUC = 0.731) and the radiomics model (AUC = 0.754) (Delong test, P < 0.05). The goodness of fit was presented by Akaike information criterion (AIC). The AIC of the radiomics–clinical model was lower than that of the clinical and radiomics models, indicating a better fitness. The DCA curve, NRI, and IDI were also used to compare the performance of various models [22]. The ROC curve compares prediction accuracy only from the standpoint of discrimination, whereas the DCA curve displays the potential risks and benefits of false negative and false positive [36]. NRI is defined as the difference in the number of correct classifications between two classifiers, which can be understood as the difference between the sums of the sensitivity and specificity of two classifiers [37]. IDI is similar to NRI in that it refers to the quantification of the prediction probability gap [38]. These two indicators are better suited for model comparison [38].

This study's ROIs were based on the arterial phase image of enhanced CT, that was, 30–35 s after the injection of enhancer, drawing lessons from Umeoka’s study. According to this study, the difference between esophageal tumor and normal esophageal wall in the second arterial phase (delayed 35 s) is significantly greater than in the first arterial phase (delayed 5 s) and venous phase (delayed 65 s) [19].

The following are the benefits of this research: for starters, this is the first radiomics–clinical prediction model for EC distant metastasis. Second, all ROIs outlined in the study included the entire esophageal tumor rather than partial tomographic images reported in previous studies, which can better present biological properties of whole tumors and have better repeatability. Third, ROIs were manually outlined to avoid the lumen area and minimize the impact of lumen contents on ROIs. Furthermore, the model was presented as a nomogram, which can more intuitively show the impact of various parameters on the outcome and is more convenient for clinical application.

The current study has some limitations. First, despite the model's strong performance, it lacked the external validation sets necessary to back up its generalization. For further validation, it is, therefore, necessary to include patients from other centers in the follow-up study. Second, because this was a retrospective study, selection bias was unavoidable, even though we used strict inclusion criteria. Third, because the goal of this study was to develop a reliable metastasis prediction model, the mechanism was not thoroughly investigated, necessitating more thorough investigations.

留言 (0)

沒有登入
gif