Prediction of a T-cell/MHC-I-based immune profile for colorectal liver metastases from CT images using ensemble learning

Ralph Saber,David Henault,Rolando Rebolledo,Simon Turcotte,Samuel Kadoury

2023-03-07

Abstract:Colorectal cancer liver metastases (CLM) are the most common type of distant metastases originating from the abdomen and are characterized by a high recurrence rate after curative resection. It has been previously reported that CLM presenting a low cluster of differentiation 3 (CD3) positive T-cell infiltration density concurrent with a high major histocompatibility complex class I (MHC-I) expression were associated with poor clinical outcomes. In this study, we attempt to noninvasively predict whether a CLM exhibit the CD3LowMHCHigh immunological profile using preoperative CT images. To this end, we propose an ensemble network combining multiple Attentive Interpretable Tabular learning (TabNet) models, trained using CT-derived radiomic features. A total of 160 CLM were included in this study and randomly divided between a training set (n=130) and a hold-out test set (n=30). The proposed model yielded good prediction performance on the test set with an accuracy of 70.0% [95% confidence interval 53.6%-86.4%] and an area under the curve of 69.4% [52.9%-85.9%]. It also outperformed other off-the-shelf machine learning models. We finally demonstrated that the predicted immune profile was associated with a shorter disease-specific survival (p = .023) and time-to-recurrence (p = .020), showing the value of assessing the immune response.

Quantitative Methods,Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to non - invasively predict whether colorectal liver metastases (CLM) have an immune signature of low CD3 - positive T - cell infiltration density and high major histocompatibility complex class I (MHC - I) expression (i.e., CD3LowMHCHigh immunophenotype) through preoperative CT images. This specific immunophenotype is associated with poorer clinical outcomes, such as shorter disease - specific survival (DSS) and time to recurrence (TTR), as well as a poorer response to chemotherapy.

Specifically, the researchers proposed an integrated network that combines multiple Attention - explainable Tabular Learning (TabNet) models and is trained using radiomic features extracted from CT images. The aim was to evaluate the predictive performance of this model on the test set and to explore whether the predicted immunophenotype could stratify patient prognoses, especially DSS and TTR. The research results show that the proposed multi - TabNet integrated model performs well in predicting the CD3LowMHCHigh immunophenotype of CLM, and the predicted immunophenotype is significantly associated with the patients' shorter DSS and TTR, indicating that CT - image - based immunophenotype prediction has potential prognostic value.