Artificial intelligence combining radiomics and clinical data for predicting response to immunotherapy

Abstract

Background There are currently no good indicators of which patients with cancer will respond or not to immunotherapy. Novel computational analysis of computed tomography scans (CT) (i.e. radiomics) provides information about the tumour-infiltrating CD8 and predict response to immunotherapy. We aim to validate in an external cohort the VHIO CT-radiomics signature and to develop a combined radiomics-clinical signature that predicts the response to immune checkpoint inhibitors in patients with advanced solid tumours. Methods The VHIO CT-radiomics signature was developed in a population of 115 consecutive patients treated with immune checkpoint inhibitors (programmed-death protein 1 [PD-1] or programmed-death ligand 1 [PD-L1] inhibitors) monotherapy in phase I clinical trials (Cohort 1). The external validation included 62 consecutive patients with urinary bladder cancer treated with anti-PD-1 or PD-L1 monotherapy (Cohort 2). From the baseline CT, a target lesion per patient was delineated. Radiomics variables of first-order, shape, and texture were extracted. An elastic-net model combining radiomics and clinical features was implemented. The association between the radiomics score and changes in tumour shrinkage was assessed using Mann-Whitney analysis. Results In the Cohort 1 the CT-radiomics signature associates with response (area under the curve [AUC] of 0.81, p-value=2.74x10-5 and 0.72, p=0.001 in the training and internal validation sets, respectively). In the external validation set (Cohort 2), the CT-radiomics signature predicts a response with an AUC of and 0.76 (p=0.001). The model combining radiomics and clinical features has an AUC of 0.84 (p-value=5.04x10-9) for response prediction. Tumour homogeneity, hypodensity and spherical shape together with high lymphocytes and albumin and low neutrophils, corresponding to a high clinical-radiomics signature score, are indicators of tumour response. A higher CT-radiomics signature score is associated with a larger tumour shrinkage (p<0.05). Conclusions