Establishing a survival prediction model for esophageal squamous cell carcinoma based on CT and histopathological images

• Published in: Physics in medicine & biology 2021-07, Vol.66 (14), p.145015
• Main Authors: Wang, Jinlong, Wu, Lei-Lei, Zhang, Yunzhe, Ma, Guowei, Lu, Yao
• Format: Article
• Language: English
• Subjects: CT radiomics; Esophageal Neoplasms - diagnostic imaging; esophageal squamous cell carcinoma (ESCC); Esophageal Squamous Cell Carcinoma - diagnostic imaging; feature extraction; histopathological image; Humans; Machine Learning; pattern recognition; prognosis; Retrospective Studies; Tomography, X-Ray Computed
• ISSN: 0031-9155; 1361-6560
• DOI: 10.1088/1361-6560/ac1020

Currently, the incidence of esophageal squamous cell carcinoma (ESCC) in China is high and its prognosis is poor. To evaluate the prognosis of patients with ESCC, we performed computerized quantitative analyses on diagnostic computed tomography (CT) and digital histopathological slices. A retrospective study was conducted to assess the prognosis of ESCC in 153 patients who underwent esophagectomy, and the cohort was selected based on strict clinical criteria. Each patient had an enhanced CT image, and there were two imaging protocols for CT images of all patients. Each patient in the cohort also had a histopathological tissue slide after hematoxylin-eosin staining. Under an electron microscope, the tissue slide was scanned as an image of large size. We then performed quantitative analyses to identify factors related to the prognosis of ESCC on digital histological images and diagnostic CT images. For CT images, we used the radiomics method. For histological images, we designed a set of quantitative features based on machine learning algorithms, such as K-means and principal component analysis. These features describe the patterns of different cell types in histopathological images. Subsequently, we used the survival analysis model established using only CT image features as the baseline. We also compared multiple machine learning models and adopted a five-fold cross-validation method to establish a robust survival model. In establishing survival models, we first used CT image features to establish survival models, and the C-index from the Weibull Cox model on the test set reached 0.624. Then we used histopathlogical features to establish survival models, and the C-index from the Weibull Cox model on the test set reached 0.664, which was obviously better than CT's. Lastly, we combined CT image features and histopathological image features to establish survival models. The performance was better than that in the models built using only CT image features or histopathological image features, and the C-index from the regularized Cox model on the test set reached 0.694. We also proved the effectiveness of the quantified histopathological image features in terms of prognosis using the log-rank test. Histopathological image features are more relevant to prognosis than features extracted from CT images using radiomics. The results of this study provide clinicians with a reference to improve the survival rate of patients with ESCC after surgery. These results have i