RADAR: An Agentic Assistant for Renal Mass Assessment and Management Planning on CT

Renal cell carcinoma (RCC) is primarily diagnosed and monitored on computed tomography (CT), where physicians must identify a tumor, characterize its features, predict aggressiveness, and decide on treatment plan (active surveillance, biopsy, thermal ablation, partial nephrectomy, or radical nephrectomy). These decisions rely on clinical judgement that integrates imaging features with a patient's history and molecular and histologic data if available. In order to support physicians in the renal mass assessment and management planning task, we introduce RADAR, a multimodal agentic assistant that integrates a large language model (LLM) with image-analysis tools and retrieval-augmented generation (RAG) designed to analyze renal masses. RADAR is designed to answer questions about renal mass assessment and management from CT and clinical data, providing the measurements and reasoning behind its answers. It is evaluated on a benchmark of 418 items constructed from 20 cases of the KiTS23 dataset and 21 annotated multiple-choice questions spanning description, surgical feasibility, biopsy, surveillance and ablation planning. RADAR with a Qwen2.5-7B LLM selects the reference answer in 80.6% of items, compared to 63.4% for GPT-4o given lesion-centered slices and 45.7% given uniformly spaced slices. RADAR is under active development to improve LLM reasoning and tool performance under multidisciplinary physician supervision spanning urology, interventional radiology, and diagnostic radiology.