Real-world evaluation of a deep learning–based serial CT imaging biomarker with variable thoracic and abdominopelvic coverage in metastatic colorectal and kidney cancers.

Abstract

e15512 Background: Early on-treatment biomarkers that predict overall survival may benefit treatment decisions and clinical trial evaluations, particularly when tumor-burden changes are weakly associated with survival, such as with immune-based therapies. Deep learning applied to routinely acquired CT may capture prognostic signal beyond tumor size, but real-world scans vary in anatomic coverage. We evaluated the real-world generalizability of a serial deep learning CT biomarker across abdominal malignancies and variable anatomic imaging. Methods: Serial CT response score (Serial CTRS) is a fully automated deep learning biomarker that predicts overall survival (OS) using baseline and early on-treatment CT scans. Serial CTRS previously demonstrated improved OS prediction compared to tumor-sized based metrics when trained on thoracic CT scans from advanced non-small cell lung cancer (aNSCLC) patients. In the present study, Serial CTRS was expanded to use thoracic, abdominal, and pelvic CT images as available, trained using real-world aNSCLC data (1,178 patients; 16,790 CT series). Serial CTRS, generalized for variable anatomical coverage, was retrospectively applied to two real world cohorts with abdominal primary cancers: 92 patients with metastatic renal cell and other kidney cancer treated with immune checkpoint inhibitors (RCC cohort) and 76 patients with metastatic colorectal cancer treated with chemotherapy (CRC cohort). Serial CTRS was generated using pretreatment scans paired with follow up scans within 28 to 120 days from treatment start. Prognostic performance of Serial CTRS was assessed using OS concordance index (C-index) and area under the receiver operating characteristic curve (AUROC) of landmark OS at 6, 12, and 24 months. Results: In the RCC cohort, Serial CTRS demonstrated a C-index of 0.76 (95% CI: 0.68-0.85), and AUROCs of 0.86 for OS6 (95% CI: 0.76-0.95), 0.82 for OS12 (95% CI: 0.70-0.93), and 0.77 for OS24 (95% CI: 0.65-0.89). Five of the 92 RCC patients did not have thoracic scan coverage. In the CRC cohort, Serial CTRS demonstrated a C-index of 0.65 (95% CI: 0.56-0.73), with AUROCs of 0.78 (95% CI: 0.59-0.96) for OS6, 0.72 (95% CI: 0.58-0.87) for OS12, and 0.67 (95% CI: 0.55-0.79) for OS24. Twenty-nine of the 76 CRC patients did not have thoracic scan coverage. Confidence intervals for all metrics excluded 0.5, consistent with non-random prognostic performance. Conclusions: Serial CTRS demonstrated prognostic associations with OS in metastatic kidney and colorectal cancer cohorts. The ability to automatically derive prognostic signal from CT imaging with variable anatomic coverage supports generalization of Serial CTRS to real-world clinical settings. Further studies, including comparisons with tumor-size based metrics, are underway to better quantify the potential utility of Serial CTRS.