Prediction of CD8+T lymphocyte infiltration levels in gastric cancer from contrast-enhanced CT and clinical factors using machine learning

Abstract

Background

CD8+ T lymphocyte infiltration is closely associated with the prognosis and immunotherapy response of gastric cancer (GC). For now, the examination of CD8 infiltration levels relies on endoscopic biopsy, which is invasive and unsuitable for longitude assessment during anti-tumor therapy.

Purpose

This work aims to develop and validate a noninvasive workflow based on contrast-enhanced CT (CECT) images to evaluate the CD8+ T-cell infiltration profiles of GC.

Methods

GC patients were retrospectively and consecutively enrolled and randomly assigned to the training (validation) or test cohort at a 7:3 ratio. All patients were binary classified into the CD8-high (infiltrated proportion ≥ 20%) or CD8-low group (infiltrated proportion < 20%) group. A total of 1170 radiomics features were extracted from each presurgical CECT series. After feature selection, fifteen radiomics features were transmitted to three independent machine-learning models for the computation of predictive radiological scores. Multilayer perceptron (MLP) was applied to merge the radiological scores with clinical factors. The predictive efficacy of the radiological scores and of the combined model was evaluated by receiver operating characteristic curve, calibration curve, and decision curve analysis in both the training and test cohorts.

Results

A total of 210 patients were enrolled in this study (mean age: 63.22 ± 8.74 years, 151 men), and were randomly assigned to the training set (n = 147) or the test set (n = 63). The merged radiological score was correlated with CD8 infiltration in both the training (p = 1.8e−10) and test cohorts (p = 0.00026). The combined model integrating the radiological scores and clinical features achieved an area under the curve (AUC) value of 0.916 (95% CI: 0.872–0.960) in the training set and 0.844 (95% CI: 0.742–0.946) in the test set for classifying CD8-high GCs. The model was well-calibrated and exhibited net benefit over “treat-all” and“treat-none” strategies in decision curve analysis.

Conclusions

Artificial intelligent systems combining radiological features and clinical factors could accurately predict CD8 infiltration levels of GC, which may benefit personalized treatment of GC in the context of immunotherapy.