Non-invasive measurement of PD-L1 status and prediction of immunotherapy response using deep learning of PET/CT images

Abstract

Background: Currently, only a fraction of patients with non-small cell lung cancer (NSCLC) treated with immune checkpoint inhibitors (ICIs) experience a durable clinical benefit (DCB). According to NCCN guidelines, Programmed death-ligand 1 (PD-L1) expression status determined by immunohistochemistry (IHC) of biopsies is the only clinically approved companion biomarker to trigger the use of ICI therapy. Based on prior work showing a relationship between quantitative imaging and gene expression, we hypothesize that quantitative imaging (radiomics) can provide an alternative surrogate for PD-L1 expression status in clinical decision support.

Methods: ¹⁸F-FDG-PET/CT images and clinical data were curated from 697 patients with NSCLC from three institutions and these were analyzed using a small-residual-convolutional-network (SResCNN) to develop a deeply learned score (DLS) to predict the PD-L1 expression status. This developed model was further used to predict DCB, progression-free survival (PFS), and overall survival (OS) in two retrospective and one prospective test cohorts of ICI-treated patients with advanced stage NSCLC.

Results: The PD-L1 DLS significantly discriminated between PD-L1 positive and negative patients (area under receiver operating characteristics curve ≥0.82 in the training, validation, and two external test cohorts). Importantly, the DLS was indistinguishable from IHC-derived PD-L1 status in predicting PFS and OS, suggesting the utility of DLS as a surrogate for IHC. A score generated by combining the DLS with clinical characteristics was able to accurately (C-indexes of 0.70-0.87) predict DCB, PFS, and OS in retrospective training, prospective testing and external validation cohorts.

Conclusion: Hence, we propose DLS as a surrogate or substitute for IHC-determined PD-L1 measurement to guide individual pretherapy decisions pending in larger prospective trials.

Keywords: immunotherapy; tumor biomarkers.

Figure 1

Study design, which contains three main phases. First, the SPH data comprised PD-L1 expression data and the corresponding imaging data was used to train and validate the deeply learned score (DLS). Then, according to FDA SaMD guideline, the DLS was evaluated through the clinical association and analytic validation on the two cohorts (MCC PD-L1 data and external VA PD-L1 data), which had both PD-L1 expression data and clinical follow-up information, as well as the clinical validation on three other cohorts (MCC ICI-treated retrospective, prospective, and external VA ICI-treated cohorts), which had clinical follow-up information. Third, in order to further test the application of DLS in guiding treatment, the well-validated DLS was utilized to develop prognosis prediction models with the MCC ICI-treated retrospective cohort, which was tested with MCC ICI-treated prospective and external VA ICI-treated cohorts. DCB, durable clinical benefit; ICI, immune checkpoint inhibitor; MCC, H Lee Moffitt Cancer Center and Research Institute; ROI, region of interest; SPH, Shanghai Pulmonary Hospital; VA, James A Haley Veterans’ Administration; EGFR, epidermal growth factor receptor; PD-L1, programmed death-ligand 1; FDA, Food and Drug Administration; PET/CT, positron emission tomography/computed tomography.

Figure 2

NSCLC histology subtypes and PD-L1 expression. Squamous cell carcinoma (SCC) patients with positive PD-L1 expression (A) and negative PD-L1 expression (B). Adenocarcinoma (ADC) patients with positive PD-L1 expression (C) and negative PD-L1 expression (D), respectively. For (A)–(D), the first line is the CT, PET, and fusion images, the first and second columns of the second and third line show the response of the fourth ResBlock, which shows the self-learned important areas in expressing PD-L1 status (peritumoral and necrosis regions), the third column of the second and third line shows the response of the negative filter and the positive filter in the PD-L1 positive–negative tumors (the CT images were overlapped to reveal the location of the response), the last line shows the pathological examination of the resected mass demonstrating PD-L1 expression (left, ×100; right, ×200). (E) The heatmap generated with unsupervised hierarchical clustering of all the SPH patients and MCC PD-L1 patients on the horizontal axis and deeply learned features expression (ie, the output of the last activation filters, N=256) on the vertical axis. There were four distinct subgroups obtained. Groups G1 and G2 (including more PD-L1− patients) had similar feature expression, which is opposite to the feature expression of G3 and G4 (including more PD-L1+ patients). Furthermore, some features of G1 and G2 (or G3 and G4) are different. G1 and G3 had more patients with SCC, while G2 and G4 had more patients with ADC. The χ² test showed the significant association of the four kinds of deep learning expression patterns with PD-L1 expression (SPH patients: p<0.001, MCC patients: p<0.001) and different histology (SPH patients: p<0.001, MCC patients: p=0.061). The similar patterns of the external MCC PD-L1 cohorts further showed the stability of the deep learning features. ADC, adenocarcinoma; PD-L1, programmed death-ligand 1; SUV, standardized uptake value; DLS, deeply learned score; MCC, H Lee Moffitt Cancer Center and Research Institute; NSCLC, non-small cell lung cancer; SCC, squamous cell carcinoma; SPH, Shanghai Pulmonary Hospital

Figure 3

Performance of the DLS in predicting PD-L1 status. (A) The distribution of DLS between PD-L1-positive (+) and PD-L1-negative (−) groups in SPH training, SPH validation, and external MCC PD-L1 test cohorts. (B) The receiver operating characteristic curves of DLS and SUVmax in SPH training, SPH validation, and external MCC PD-L1 test cohorts. (C) The quantitative performance metrics in SPH training, SPH validation, external MCC PD-L1 test, and external VA test cohorts. ACC, accuracy; AUC, area under receiver operating characteristics curve; DLS, deeply learned score; PD-L1, programmed death-ligand 1; SUV, standardized uptake value; MCC, H Lee Moffitt Cancer Center and Research Institute; SEN, sensitivity; SPEC, specificity; SPH, Shanghai Pulmonary Hospital; VA, James A Haley Veterans’ Administration.

Figure 4

Performance of the DLS in prognosis prediction. (A) The ROC curve of DLS in DCB prediction, and the PFS and OS relative to the DLS (DLS cutoff: 0.55) in the retrospective MCC ICI-treated patients. (B) The ROC curve of DLS in DCB prediction, and the PFS and OS relative to the DLS (DLS cutoff: 0.55) in the prospective MCC ICI-treated patients. (C) The ROC curve of DLS in DCB prediction, and the PFS and OS relative to the DLS (DLS cutoff: 0.55) in the external VA test patients. P value was from log rank test. AUC, area under the receiver operating characteristic curve; DCB, durable clinical benefit; DLS, deeply learned score; HDLS, high DLS; ICI, immune checkpoint inhibitor; LDLS, low DLS; MCC, H Lee Moffitt Cancer Center and Research Institute; OS, overall survival; PFS, progression-freesurvival; ROC, receiver operating characteristic; VA, James A Haley Veterans’ Administration.

Figure 5

Stratification analysis of the performance of the DLS in prognosis prediction. (A) The DCB rates of the different subgroups of the MCC retrospective and prospective ICI-treated patients. (B) The PFS relative to the DLS and histology in the MCC retrospective and prospective ICI-treated patients. (C) The OS relative to the DLS and histology in the MCC retrospective and prospective ICI-treated patients. Note: HADC is short for HDLS ADC, meaning ADC patients with high DLS; LADC is short for LDLS ADC, meaning ADC patients with low DLS; HSCC is short for HDLS SCC, meaning SCC patients with high DLS; and LSCC is short for LDLS SCC, meaning SCC patients with low DLS, the high DLS versus low DLS defined by 0.55. P value was from log rank test. ADC, adenocarcinoma; DCB, durable clinical benefit; DLS, deeply learned score; ICI, immune checkpoint inhibitor; MCC, H Lee Moffitt Cancer Center and Research Institute; OS, overall survival; PFS, progression-freesurvival; SCC, squamouscell carcinoma.