Prediction of CD3 T cells and CD8 T cells expression levels in non-small cell lung cancer based on radiomic features of CT images

Abstract

Background: In this work, radiomics characteristics based on CT scans were used to build a model for preoperative evaluation of CD3 and CD8 T cells expression levels in patients with non-small cell lung cancer (NSCLC).

Methods: Two radiomics models for evaluating tumor-infiltrating CD3 and CD8 T cells were created and validated using computed tomography (CT) images and pathology information from NSCLC patients. From January 2020 to December 2021, 105 NSCLC patients with surgical and histological confirmation underwent this retrospective analysis. Immunohistochemistry (IHC) was used to determine CD3 and CD8 T cells expression, and all patients were classified into groups with high and low CD3 T cells expression and high and low CD8 T cells expression. The CT area of interest had 1316 radiomic characteristics that were retrieved. The minimal absolute shrinkage and selection operator (Lasso) technique was used to choose components from the IHC data, and two radiomics models based on CD3 and CD8 T cells abundance were created. Receiver operating characteristic (ROC), calibration curve, and decision curve analyses were used to examine the models' ability to discriminate and their clinical relevance (DCA).

Results: A CD3 T cells radiomics model with 10 radiological characteristics and a CD8 T cells radiomics model with 6 radiological features that we created both demonstrated strong discrimination in the training and validation cohorts. The CD3 radiomics model has an area under the curve (AUC) of 0.943 (95% CI 0.886-1), sensitivities, specificities, and accuracy of 96%, 89%, and 93%, respectively, in the validation cohort. The AUC of the CD8 radiomics model was 0.837 (95% CI 0.745-0.930) in the validation cohort, with sensitivity, specificity, and accuracy values of 70%, 93%, and 80%, respectively. Patients with high levels of CD3 and CD8 expression had better radiographic results than patients with low levels of expression in both cohorts (p<0.05). Both radiomic models were therapeutically useful, as demonstrated by DCA.

Conclusions: When making judgments on therapeutic immunotherapy, CT-based radiomic models can be utilized as a non-invasive way to evaluate the expression of tumor-infiltrating CD3 and CD8 T cells in NSCLC patients.

Keywords: CD3; CD8; model; non-small-cell lung cancer; radiomics.

Figure 1

Representative immunohistochemistry staining images of CD3 and CD8 cells from NSCLC patients.

Figure 2

The workflow of our methodology. (A) To rebuild the 3D-VOI using ITK-snap, the ROI was manually formed on the lung window sequence of the CT images. (B) AK retrieved 1316 features from CT scans. Features were chosen by LOSSO and ANOVA-KW (C). (D, E) ROC, calibration curve, and DCA were used to examine the diagnostic effectiveness of the model.

Figure 3

The most advantageous subset of radiomics characteristics was extracted using the LASSO technique and 10-fold cross-validation. The best feature selection based on AUC values is shown in (A, B). The best value that yields the lowest binomial deviation is shown by the vertical dashed line’s log() value. When the ln () value rises to this level, the AUC hits a peak matching to the ideal number of radiomics characteristics. (C, D) Radiomics portion LASSO coefficient distribution.

Figure 4

Heat map showing correlations for identifying histological features in CD3 (A) and CD8 (B) expression imaging. The values of the color bars on the right are the correlation coefficients, and hues imply strong and positive correlations while light colors suggest negative correlations.

Figure 5

Radiomic scores of patients in different cohorts. The radiomic scores of patients with strong infiltration of CD3 and CD8 cells were substantially higher than those of patients with low infiltration in both the training cohort and validation cohort.

Figure 6

Adiomics score (Rad-score) bar graph for the CD3 (A) and CD8 (B) training set. The Rad-score value is shown on the Y-axis; positive values indicate high expression forecasts, negative values indicate low expression predictions, red bars indicate true soft expression cases, and blue bars indicate real high expression cases. Correct predictions are represented by red bars with negative values and blue bars with positive values, whereas wrong predictions are blue bars with negative values and red bars with positive values.

Figure 7

(A, B) ROC curves of the CT texture feature model predicting CD3 expression. a ROC curve of the training cohort (n=72). (B) shows ROC curves of the validation cohort (n=33) (C, D) ROC curves of CD8 expression predicted by CT texture feature model. ROC curves of the training cohort (n=73). (B) shows ROC curves of the validation cohort (n=32).

Figure 8

Column line plot calibration curves for the CD3 training set (A) and validation set (B). The incidence of projected CD3 expression is represented by the horizontal axis. The incidence of detected CD3 expression makes up the vertical axis. The reference line, which is red on the diagonal, shows that the predicted value and the actual value are same. The fact that the forecasts mainly coincide with the diagonal line shows that they are correct. Column line plot calibration curves for the CD8 training set (C) and validation set (D). The projected incidence of CD8 expression is represented on the horizontal axis. The incidence of detected CD8 expression makes up the vertical axis. The reference line, which is red on the diagonal, shows that the anticipated value and the actual value are same. The fact that the expected outcomes generally coincide with the diagonal line shows that the predictions were correct.

Figure 9

Non-small cell lung cancer (NSCLC) CD3 (A) and CD8 (B) expression is predicted using a decision curve analysis, with the horizontal axis displaying the range of risk thresholds and the vertical axis displaying the net benefit. The gray line in the illustration denotes the presumption that all lesions are positively expressed, and the more the blue curve deviates from the gray line, the greater the net advantage of the model.