Abstract

Contrast-enhanced computed tomography (CT) is central to the diagnosis, staging, and follow-up of patients with renal cell carcinoma (RCC). As artificial intelligence research into computer-aided solutions continues to grow, the need for curated and annotated datasets becomes increasingly important. Imaging-based artificial intelligence studies often need lesion annotations that are not consistently available. The Cancer Genome Atlas (TCGA) datasets are widely used for model training and validation. However, access to public annotations of lesions is limited, which limits reproducibility and comparability of the published research. To address this gap, we screened 1,915 CT scans from three TCGA-RCC databases and, following a meta-data-based exclusion step, used an automated segmentation model to generate initial kidney and lesion masks. Next, we conducted a reader study with all papillary (n=56), chromophobe (n=27) and 200 randomly selected clear cell RCC cases. Two trained students performed quality checks, corrections, and additional annotation of tumors and cysts, with uncertain cases reviewed by a board-certified radiologist. After data exclusion and quality control, a final cohort of 129 annotated CT scans from 91 patients (24 female, 67 male; mean age 56 years) was retained, including 85 clear cell, 26 papillary and 18 chromophobe RCC cases. Images and voxel-level annotations of kidneys and lesions are openly available at https://zenodo.org/records/20719257. By open-sourcing these annotations, we aim to foster accessible, reproducible AI research in renal cell carcinoma. RCC-AID provides a reusable open resource dataset for segmentation, detection, subtype classification, radiomics, and multimodal RCC research.