Kidney cancer screening in non-contrast computed tomography via domain-adaptive deep learning

Early detection of kidney cancer improves prognosis, but screening is hindered by the lack of reliable, cost-effective biomarkers. Imaging paired with Artificial Intelligence (AI) is a highly promising avenue for screening. Here we present an AI framework for kidney cancer screening via low-dose and full-dose non-contrast computed tomography (NCCT). We developed a two-stage deep learning framework for renal mass detection, trained on over 1,300 multi-institutional computed tomography (CT) scans using the novel RAdiological Domain-Adaptive Recognition (RADAR) paradigm. The RADAR pre-training method leverages spatial correspondences in unlabelled multiphasic CT data to improve lesion detection accuracy. Our model achieved an area under the curve (AUC) of 0.938 on low-dose NCCT in the internal validation set, outperforming all five radiologists in a reader study. We externally validated RADAR in 3,999 participants from the NCCT-based Yorkshire Kidney Screening Trial (YKST). Our model detected 100% of the histologically confirmed kidney cancers found in YKST with 93.4% specificity, achieving an AUC of 0.980, while also identifying seven of eight renal masses without histological confirmation, flagging them for further diagnostic assessment. Health economic simulations suggest that AI integration could improve early detection while reducing screening costs by 44% and radiation exposure by 29%.