Improved Immunotherapy Response Prediction in NSCLC With Deep-Learning Radiomic Biomarker

Use of a fully automated deep learning radiomic biomarker based on serial CT scans resulted in more effective overall survival predictions in patients with advanced non–small cell lung cancer (NSCLC) receiving immune checkpoint inhibitors than resource-intensive Response Evaluation Criteria in Solid Tumors (RECIST) and tumor volume change measurements.

“These findings suggest that Serial CT response scores could improve clinical decision-making and enhance clinical trial designs for patients with NSCLC,” Sako et al commented in the report of their prognostic study published in JAMA Network Open.

Study Details

The investigators analyzed retrospectively collected electronic health record data from routine clinical practice and clinical trial data from 2013 to 2023.

The study included 1,830 patients with advanced NSCLC who initiated immune checkpoint inhibitor therapy between 2013 and 2021 in the discovery cohort (n = 1,171), 2013 and 2022 in the test cohort (n = 605), and 2017 and 2018 in the GARNET cohort (n = 54). The study population had a median age of 67 years and included 1,000 males (55%) and 830 females (45%).

Cox proportional hazards regression and receiver operating characteristic area under the curve analyses were used to model associations between the Serial CTRS and overall survival.

Model Methods

The Serial CTRS model was developed using a multistage pipeline of 14,424 unannotated thoracic 3D CT images taken from pretherapy and for 12-week follow-ups from 1,171 patients that were retrospectively collected between 2020 and 2024. CT images were taken using different scanner models and acquisition/reconstruction settings.

The images were preprocessed and input into a feature extractor along with paired lesion autosegmentation masks. Deep learning techniques focused the feature extraction on areas associated with advanced NSCLC.

Overall, the AI model for generating Serial CTRS consisted of a custom 3D, hierarchical, convolutional neural network for multiple survival endpoints, with dropout and augmentation techniques added to limit model overfitting. The determined imaging features were then put into Cox proportional hazard models that were trained on overall survival assessments, and cross-validation further optimized the model output. Serial CTRS then could generate a continuous score between 0 and 1 to represent a probability of overall survival at 1 year.

The model was then validated with routine clinical practice test datasets from 10 institutions in the United States and Europe, and independently validated on the multinational phase I GARNET trial of dostarlimab-gxly.

Key Findings

Based on multivariable analysis, after controlling for age, sex, PD-L1 expression, histologic profile, and tumor volume, the Serial CTRS was associated with overall survival (hazard ratio [HR] per 10-percentage-point increase in predicted 12-month overall survival: 0.74, 95% confidence interval [CI] = 0.70–0.79 in the test cohort; 0.45, 95% CI = 0.32–0.65 in the GARNET cohort). The Serial CTRS outperformed RECIST and tumor volume change in discriminating overall survival risk, the investigators noted, with higher HRs distinguishing low- and high-survival groups in both the test (HR = 6.19, 95% CI = 4.12–9.28) and GARNET (HR = 18.00, 95% CI = 5.40–59.97) cohorts. The biomarker appeared to maintain its predictive value across PD-L1 and RECIST subgroups, including stable disease.

The investigators concluded, “Serial CTRS is an externally validated, fully automated, deep-learning, serial imaging–based biomarker that leverages routine CT scans from baseline and early-response follow-up to predict overall survival more effectively than RECIST and tumor volume change in immune checkpoint inhibitor–treated patients with advanced NSCLC. The automated design of Serial CTRS facilitates future integration into clinical practice and clinical trial workflows. With further validation across therapeutic modalities, Serial CTRS has the potential to enable more accurate, early treatment readouts in both clinical practice and clinical trial settings.”

Chiharu Sako, PhD, of Onc.AI, San Carlos, California, is the corresponding author of the article in JAMA Network Open.

Disclosure: The GARNET study was funded by GSK. This project was funded in part by the National Institutes of Health. Dr. Sako reported employment and stock options with Onc.AI. For full disclosures of all study authors, visit jamanetwork.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.