OPTimizing surveillance in lung cancer survivors with novel IMAging biomarkers and deep-Learning (OPTIMAL)

ABSTRACT Over the next decade, the number of lung cancer survivors is expected to rise considerably in the United States, due to an aging and growing population and a shift toward more early-stage lung cancers detected from increased lung cancer screening uptake. After treatment with curative-intent surgery, however, many survivors of early-stage non-small cell lung cancer (NSCLC) remain at high risk for lung cancer recurrence or second primary lung cancer (SPLC). To detect recurrence and SPLC at their earliest stages, postoperative imaging surveillance with chest computed tomography (CT) is generally recommended every six months for the first two years and annually thereafter. Yet, this clinical recommendation is based on limited evidence of benefits and harms, with no consideration for individual differences in risk of recurrence or SPLC. Routine chest CT images are a potentially useful, but largely untapped, data source for predicting risk of recurrence or SPLC to tailor surveillance among lung cancer survivors. Chest CT scans contain specific imaging biomarkers of body composition and cardiopulmonary health, as well as non-specific imaging data that are amenable to analysis using deep learning, all ascertainable without additional intervention, risks, or costs to survivors. Studies including ours suggest that skeletal muscle and adipose tissue measured from CT scans can predict outcomes after NSCLC resection. In the context of lung cancer screening, we have also developed Sybil, a validated deep- learning model that uses information from a single low-dose chest CT scan, to accurately predict future incident lung cancer risk. Our overarching goal is to optimize survivorship of early-stage NSCLC following curative-intent surgery by incorporating a risk-based surveillance strategy that leverages routinely available imaging data. Our multidisciplinary team proposes to examine 12,000 individuals treated surgically for stage I or II NSCLC from 2015 to 2025, with follow-up for outcomes through 2027, using longitudinal electronic health records and serial chest CT scans from three healthcare systems that serve distinct and diverse populations. First, we will determine real-world practice patterns and effectiveness of imaging surveillance, overall and by individual characteristics. Second, we will develop and validate risk prediction models for lung cancer recurrence and SPLC that incorporate imaging biomarkers of body composition and of vascular and pulmonary health derived from preoperative and postoperative surveillance CT scans. Lastly, we will assess and then optimize Sybil’s performance in predicting lung cancer recurrence or SPLC using postoperative surveillance CT scans. Overall, the proposed study will advance knowledge about the promising potential for personalized risk-stratified surveillance of recurrent or new disease using novel imaging-driven methods among the growing population of lung cancer survivors.