Different MRI-based radiomics machine learning models to predict CD3+ Tumor-infiltrating Lymphocytes (TILs) in rectal cancer

Weili Ma ¹ Chuanling Hou ¹ Minxia Yang ¹ Yuguo Wei ² Jiwei Mao ¹ Le Guan ¹ Zhenhua Zhao ^3*

• ¹ Shaoxing People's Hospital, Shaoxing, China
• ² Advanced Analytics, Global Medical Service, GE Healthcare, hangzhou, China
• ³ Shaoxing Maternal and Child Health Hospital, shaoxing, China

Objectives: This study aims to develop and evaluate multiple machine learning models utilizing contrast-enhanced T1-weighted imaging (T1-CE) to differentiate between low-/high-infiltration of total T lymphocytes (CD3) in patients with rectal cancer.We retrospectively selected 157 patients (103 men, 54 women) with pathologically confirmed rectal cancer diagnosed between March 2015 and October 2019. The cohort were randomly divided into training dataset (n=109) and test dataset (n=48) for subsequent analysis. Seven radiomic features were selected to generate three models: LR, RF, and SVM. Diagnostic performance of the three models was compared using the Delong test. Additionally, Kaplan-Meier analysis was employed to assess DFS in patients with high and low CD3+ TIL density.The three radiomics models perform well in predicting the infiltration of CD3+ TILS, with AUC values of 0.871, 0.982, and 0.913, respectively, in the training set for the LR, RF, and SVM models. In the validation set, the corresponding AUC values were 0.869, 0.794, and 0.837, respectively. Among the radiomics models, the LR model exhibited superior diagnostic performance and robustness. The merged model, which integrated radiomics features from SVM model and clinical features from clinical model, outperformed the individual radiomics models, with AUCs of 0.8932 and 0.8829 in the training and test cohorts, respectively. Additionally, a lower expression level of CD3+ TILs cohort was independently correlated with disease-free survival (DFS) (P = 0.0041).The combined model demonstrated higher discriminatory in assessing the abundance of CD3+ TILs in rectal cancer. Furthermore, the expression of CD3+ TILs was significantly correlated with DFS, highlighting its potential prognostic value.This study is the first attempt to compare the predictive TILs performance of three machine learning models, logistic regression (LR), random forests (RF), and support vector machine (SVM) models, based on the combination of radiomics and immunohistochemistry. MRI-based combined model, composed of radiomics features from SVM model and clinical features from clinical model, exhibited higher discriminatory capability of an abundance of CD3+ TILs in rectal cancer.