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

Abstract: AI has achieved remarkable success in medical imaging, but it is widely recognized that these models often perform inconsistently across real-world clinical settings. Such inconsistencies occur when patient demographics and imaging protocols vary, for example, in detecting small tumors, analyzing scans from different contrast phases, or evaluating patients of different ages or sexes. To quantify these inconsistencies, we develop a large-scale, open benchmark of 85,355 CT scans that systematically evaluates 12 tumor-detection AI models across tumor size, location, patient subgroup, and imaging protocol. We leverage LLMs to extract and organize subgroup information from clinical data, which makes the analysis both scalable and reproducible. Our benchmark reveals that current state-of-the-art AI models, optimized for average accuracy, perform poorly in rare or underrepresented subgroups, such as young, female African Americans. However, collecting sufficient annotated data for these rare cases is often impractical. The benchmark provides a foundation for building more reliable and robust AI models for tumor detection and highlighting the need for rigorous, subgroup-level evaluation in medical imaging and computer vision. Datasets, code

• The paper reveals that common tumor detection models drastically underperform in minority and protocol-biased subgroups.
• It leverages a large, diverse CT scan dataset of 85,355 images with rigorous metadata extraction and radiologist adjudication.
• The benchmark evaluates 12 3D segmentation architectures and shows synthetic data augmentation partially mitigates bias, urging subgroup-aware evaluations.

Comprehensive Benchmarking of AI for Cancer Detection: Demographic and Protocol Biases in BenchX

Motivation and Problem Formulation

BenchX addresses a critical challenge in medical imaging AI: quantifying and mitigating demographic and protocol bias in cancer detection. While advances in deep learning have driven substantial improvements in tumor segmentation and detection on homogeneous datasets, the failure modes arising from demographic skew (age, sex, race) and imaging protocol heterogeneity (contrast phase) are poorly characterized. Existing benchmarks—MSD, FLARE, PANORAMA—overemphasize mean accuracy, masking subgroup collapses that are clinically consequential. BenchX introduces a large-scale, subgroup-aware benchmark to systematically expose and quantify these fairness and domain generalization gaps in tumor detection and localization models (2606.24883).

Dataset Construction and Metadata Extraction

BenchX compiles 85,355 abdominal CT scans from six geographically diverse hospital cohorts. The dataset spans broad demographic, racial, and age distributions, as well as multiple CT imaging protocols, including arterial, venous, non-contrast, and delayed phases. Tumor annotation combines high-sensitivity segmentation and independent scan-level detection models; discordant cases are resolved with radiologist adjudication to ensure robust labeling. Crucially, subgroup metadata—age, sex, race, CT phase, scanner type, voxel spacing—are extracted at scale via LLMs, employing strict schema normalization and evidence quoting to maximize reproducibility and minimize noise.

Quality control procedures involve LLM cross-validation and targeted human audits, achieving >98% agreement for core attributes. Missing or ambiguous information is explicitly labeled as "Unknown" to prevent spurious subgroup analysis. This structured, scalable pipeline enables fine-grained performance assessment across demographic and protocol strata.

Benchmark Analysis and Model Evaluation

BenchX evaluates 12 prominent 3D segmentation architectures: classical CNNs (U-Net, Universal Models), Transformer-based encoders (UNETR, Swin UNETR variants, VSmTrans), scalable CNNs (STU-Net-B, MedNeXt), and automated pipelines (nnU-Net, ResEnc). Models are trained on AbdomenAtlas 2.0 (pancreatic tumors + controls), following Touchstone-validated hyperparameters to enforce fair comparisons.

Evaluation metrics include sensitivity, specificity, F1-score, and balanced AUC—metrics essential for dissecting errors under population skew and protocol shifts. Data splits are cohort-exclusive, with patient-level randomization; subgroup statistics are computed only for cases with definitive labels.

Subgroup-Level Performance Degradation

BenchX reveals systemic vulnerabilities in state-of-the-art tumor detection models:

• Demographic Bias: Performance on young patients is severely degraded. For Age ≤ 20, multiple models collapse to F1 = 0, driven by low sensitivity (missed positives). Disparities persist in racial subgroups—Black and Pacific Islander patients, consistently underrepresented, yield F1 scores near zero across most architectures.
• Protocol Bias: Imaging protocol is a dominant factor, rivaling architectural choice. Venous phase scans yield the highest sensitivity and F1; arterial and non-contrast scans cause 15–30 point reductions, with multiple models exhibiting near-zero sensitivity on non-contrast data.
• Architectural Robustness: MedNeXt exhibits the most stable performance overall but fails catastrophically in certain subgroups. No architecture achieves top performance across all demographic or protocol strata, underscoring the inadequacy of average accuracy as a measure of real-world robustness.
• Error Modes: In hard subgroups, specificity remains high, but sensitivity (recall) collapses, indicating mostly missed tumors rather than false positives.

These findings are consistent across six external cohorts and multiple evaluation metrics.

Synthetic Data and Domain Adaptation

BenchX tests conditional diffusion-based synthetic data augmentation (SynthX), generating tumors conditioned on demographic and protocol attributes. Adding synthetic data significantly increases subgroup sensitivity and average F1 (e.g., +22.0% for Age 60–80, +7.8% for White), demonstrating partial mitigation of bias. However, gaps remain, particularly in underrepresented racial and age strata, implying that synthetic augmentation alone is insufficient.

Theoretical and Practical Implications

BenchX establishes that medical imaging AI, when evaluated via average metrics, fails to capture catastrophic subgroup losses. These failures stem from class imbalance, protocol heterogeneity, and insufficient demographic representation. The benchmark exposes critical limitations in "fairness-aware" medical AI: standard metrics and training protocols are ill-equipped for real-world deployment where subgroup skew is intrinsic.

For AI system safety and regulatory approval, subgroup-aware evaluation is mandatory. Practical implications include the necessity for continual data diversification, protocol harmonization, and model selection validation on external datasets. Theoretical implications extend to algorithmic approaches for robustness and fairness, including weighted loss functions, domain adaptation, meta-learning, and synthetic data strategies—each needing systematic validation against multi-cohort benchmarks like BenchX.

Future Directions

The benchmark motivates several lines of inquiry:

• Fairness-aware Training: Optimizing for worst-case subgroup performance rather than mean accuracy.
• Domain Generalization: Leveraging transfer learning, multi-source adaptation, and meta-learning for robust generalization.
• Synthetic Data Generation: Conditioning generative models on granular subgroups and expanding to rare pathologies.
• Automated Model Selection: Incorporating uncertainty estimation and subgroup-specific risk modeling into the deployment pipeline.

Expansion of BenchX to include foundation models, cross-modal (MRI, PET), and multi-organ datasets will enable comprehensive assessment of generalizability and transferability in medical AI.

Conclusion

BenchX provides a rigorous foundation for evaluating and improving the robustness of AI models for cancer detection across diverse populations and protocols. Current models, optimized for global accuracy, can appear competitive while catastrophically underperforming in rare or minority subgroups. Synthetic augmentation offers partial remediation but does not eliminate bias. Subgroup-aware benchmarking must become standard practice for medical imaging AI, ensuring equitable performance, mitigating risk, and supporting clinical translation (2606.24883).