CT Based Scalable Medical Artificial Intelligence for Recalcitrant Cancers Screening and Early Detection - Project Summary Pancreatic, hepatobiliary (liver/gallbladder), and upper GI (stomach, duodenum, esophagus) cancers are often found late, when cure is unlikely, and together cause more than 113,000 deaths annually in the United States. There is no broad population screening for these diseases, yet CT imaging is already part of routine care—more than 85 million scans per year—creating a practical opportunity for earlier cancer detection without additional imaging visits. Our vision is to turn these routine scans into an early-warning safety net that identifies people at high risk months before clinical recognition, so clinicians can act sooner. We will develop AI algorithms to detect very small cancers (<2 cm) in the pancreas, hepatobiliary, and upper-GI tract from two complementary types of CT scans: indicated scans (ordered for symptoms, abnormal labs, or known high risk) and opportunistic scans (acquired for other reasons, e.g., trauma evaluation, often without contrast enhancement). Our approach is designed to maximize clinical value in both contexts. For indicated scans, the goal is high sensitivity and specificity, minimizing missed detections and false alerts that could lead to unnecessary invasive follow-up such as biopsy or surgery. For opportunistic scans, where the consequence of a false alert is typically a non-invasive follow-up imaging (e.g., contrast CT or MRI) at low risk, we emphasize very high sensitivity with acceptable specificity to capture subtle or overlooked early cues that may precede visible tumor formation. To this end, we will pursue three specific aims. Aim 1 (Indicated scans): Develop AI to detect and localize tumors (diameter <2 cm, T1 stage) using hybrid supervision—combining per-voxel annotations, report-derived weak labels, synthetic tumors, and imaging biomarkers such as duct dilation and textural changes. Aim 2 (Opportunistic scans): Extend detection to non-contrast CTs via domain adaptation and paired contrast/non-contrast translation to identify early cues such as fat infiltration or atrophy (“pancreas index”) and analyze longitudinal changes. Aim 3 (Clinical validation): Perform silent retrospective and multi-reader studies using de-identified CT exams and simulated clinical viewers to assess sensitivity, specificity, lead time to detection, radiologist performance, and workflow gains from AI-drafted reports. Our team unites AI and medical-imaging experts at JHU, Stanford, and UCSF. We will leverage >1,000,000 CT scans, paired with radiology/pathology reports and electronic health record (EHR) from these institutions and public datasets that include scans from 145 hospitals worldwide, supporting diversity in patient demographics and imaging protocols. Primary endpoints are patient-level sensitivity/specificity for small tumors and PPV at clinically appropriate thresholds; secondary endpoints include reader-study improvements and workflow efficiency from AI-drafted reports. The expected outcome is a validated, deployable AI system that repurposes existing CT—both indicated and opportunistic—for earlier cancer detection in three exemplar organ systems, with methods designed to extend to multiple recalcitrant cancers that are exceptionally lethal and resistant to current treatments.
