Prediction of response to immune checkpoint inhibitors in solid tumours using CT-based biomarkers

Abstract

Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment by harnessing the immune system’s ability to target cancer cells. However, only a subset of patients clinically benefit from ICIs, highlighting the need for predictive biomarkers. Currently, FDA-approved biomarkers such as PD-L1 expression and mismatch repair deficiency have limited efficacy and require invasive tumour biopsies. An alternative approach involves the use of radiomics, which leverages quantitative analysis of medical images to extract a large number of imaging features. Unlike biopsies, radiomics analysis is non-invasive and provides insights into tumour heterogeneity at a whole-tumour level. In this study, we aimed to predict clinical benefit in patients treated with ICI therapy using radiomic features extracted from baseline Computed Tomography (CT) images. We analysed a data set of 447 patients with 13 different primary tumour types. Five aggregation methods were employed to combine features from lesion level to patient level. The so-called radiomics standard pipeline, LASSO and logistic regression was used for feature selection and classification. Additionally, we explored the impact of primary tumour location and developed tumour-specific models. The best performance was achieved in the case of bladder cancer (n = 53, AUC: 0.717) when using the all lesions per patient for feature aggregation, using the largest lesion as feature aggregation yielded better results for the rest of the analysed cohorts: the whole cohort (n = 447, AUC: 0.634), thoracic cancer (n = 108, AUC: 0.741), skin cancer (n = 79, AUC: 0.766), and lower gastrointestinal cancer (n = 64, AUC: 0.794). Interestingly, better results were obtained when using tumour-specific models. These results may indicate the importance of distinguishing between different tumour types when predicting response to ICIs. In order to enhance the accuracy of predicting responses to ICIs, future research should focus on investigating tumour-specific strategies, examining the potential benefits of incorporating additional clinical, genomics, and immunohistochemistry data and the use of deep learning techniques in larger, more representative cohorts.