Radiomics Model for Predicting TP53 Status Using CT and Machine Learning Approach in Laryngeal Squamous Cell Carcinoma

Ruxian Tian; Yumei Li; Chuanliang Jia; Yakui Mou; Haicheng Zhang; Xinxin Wu; Jingjing Li; Guohua Yu; Ning Mao; Xicheng Song

doi:10.3389/fonc.2022.823428

Radiomics Model for Predicting TP53 Status Using CT and Machine Learning Approach in Laryngeal Squamous Cell Carcinoma

Front Oncol. 2022 Apr 28:12:823428. doi: 10.3389/fonc.2022.823428. eCollection 2022.

Authors

Ruxian Tian¹, Yumei Li¹, Chuanliang Jia¹, Yakui Mou¹, Haicheng Zhang², Xinxin Wu¹, Jingjing Li¹, Guohua Yu³, Ning Mao², Xicheng Song^{1

4}

Affiliations

¹ Department of Otorhinolaryngology, Head and Neck Surgery, The Affiliated Yantai Yuhuangding Hospital of Qingdao University, Yantai, China.
² Department of Radiology, Yantai Yuhuangding Hospital, Qingdao University, Yantai, China.
³ Department of Pathology, Yantai Yuhuangding Hospital, Qingdao University, Yantai, China.
⁴ Shandong Provincial Clinical Research Center for Otorhinolaryngologic Diseases, Yantai, China.

Abstract

Objective: We aim to establish and validate computed tomography (CT)-based radiomics model for predicting TP53 status in patients with laryngeal squamous cell carcinoma (LSCC).

Methods: We divided all patients into a training set 1 (n=66) and a testing set 1 (n=30) to establish and validate radiomics model to predict TP53. Radiomics features were selected by analysis of variance (ANOVA) and the least absolute shrinkage and selection operator (Lasso) regression analysis. Five radiomics models were established by using K-Nearest Neighbor, logistics regressive, linear-support vector machine (SVM), gaussian-SVM, and polynomial-SVM in training set 1. We also divided all patients into a training set 2 and a testing set 2 according to different CT equipment to establish and evaluate the stability of the radiomics models.

Results: After ANOVA and subsequent Lasso regression analysis, 22 radiomics features were selected to build the radiomics model in training set 1. The radiomics model based on linear-SVM has the best predictive performance of the five models, and the area under the receiver operating characteristic curve in training set 1 and testing set 1 were 0.831(95% confidence interval [CI] 0.692-0.970) and 0.797(95% CI 0.632-0.957) respectively. The specificity, sensitivity, and accuracy were 0.971(95% CI 0.834-0.999), 0.714(95% CI 0.535-0.848), and 0.843(95% CI 0.657-0.928) in training set 1 and 0.750(95% CI 0.500-0.938), 0.786(95% CI 0.571-1.000), and 0.667(95% CI 0.467-0.720) in testing set 1, respectively. In addition, the radiomics model also achieved stable prediction results even in different CT equipment. Decision curve analysis showed that the radiomics model for predicting TP53 status could benefit LSCC patients.

Conclusion: We developed and validated a relatively optimal radiomics model for TP53 status prediction by trying five different machine learning methods in patients with LSCC. It shown great potential of radiomics features for predicting TP53 status preoperatively and guiding clinical treatment.

Keywords: TP53; computed tomography; laryngeal squamous cell carcinoma; machine learning; radiomics.