BenchX: Benchmarking AI Models for Cancer Detection and Localization with Demographic and Protocol Biases

A new open benchmark called BenchX, built from 85,355 CT scans, reveals that state-of-the-art AI models for tumor detection perform poorly on rare patient subgroups, including young, female African Americans, according to a preprint posted to arXiv [1]. The benchmark systematically evaluates 12 tumor-detection AI models across tumor size, location, patient subgroup, and imaging protocol [1]. The research, submitted to the arXiv preprint repository on June 23, 2026, aims to quantify inconsistencies that arise when patient demographics and imaging protocols vary in real-world clinical settings [1]. arXiv, founded in 1991, is an open-access repository that hosts scientific papers in fields including computer science and quantitative biology and has grown to receive roughly 24,000 submissions per month as of late 2024 [6]. The study leveraged large language models to extract and organize subgroup information from clinical data, a method the authors describe as both scalable and reproducible [1]. Large language models are machine learning systems trained on vast amounts of text for tasks such as language generation [8]. The use of such models allowed the researchers to structure the demographic and protocol data needed for the benchmark without manual annotation [1]. The findings show that models optimized for average accuracy falter when confronted with underrepresented populations [1]. The authors note that collecting sufficient annotated data for these rare cases is often impractical, which makes subgroup-level evaluation critical for building more reliable medical AI [1]. The benchmark’s dataset and code have been made openly available to provide a foundation for future work on robust tumor-detection models [1].