Mechanistic audit locates spurious annotation-reading attention heads in medical vision transformers

By applying mechanistic interpretability to chest X-ray classifiers, researchers identified dedicated attention heads that latch onto corner text rather than lung pathology, a bias completely invisible to standard output ablation.

Primary source: arXiv:2610.21638: https://arxiv.org/abs/2610.21638

What’s new: Vision Transformers (ViTs) applied to medical imaging are known to exploit spurious correlations, but diagnosing and repairing the underlying computational circuits has proven exceptionally difficult. In Spurious Attention Heads in Vision Transformers: Mechanistic Identification of Annotation-Driven Bias in Medical Chest X-Ray Analysis (arXiv:2610.21638, 15 Oct 2026), researchers Joshua Muhumuza and Cornelius Kalenzi apply mechanistic interpretability to localize shortcut learning in ViT-Base models trained on chest radiographs. They discover a dedicated attention head (L8H10) that consistently reads corner text annotations instead of anatomical structures, attending to text markers rather than lung pathology across 62% of 200 COVID-19 radiographs.

Causal validation using inpainting confirms that text semantics, rather than luminance or local edge density, drive the attention pattern. Removing the text annotations increases attention on actual lung fields by 1.52 times in affected scans (p < 0.0001). This phenomenon is not isolated to a single model: spurious annotation-reading heads were observed across 79% of scans in NIH ChestX-ray14 and replicated across DeiT-Base (88%) and ViT-Large (80%). Crucially, standard full fine-tuning failed to eliminate the shortcut; 98.4% of affected scans retained spurious head attention, with the overall spurious rate increasing slightly. Furthermore, behavioral evaluations such as single-head and multi-head ablation showed no statistically significant change in prediction accuracy (p = 0.340 and p = 0.532, respectively), demonstrating that output-level benchmarking completely fails to detect this circuit-level pathology.

Why it matters: Hospital datasets frequently feature peripheral text labels, radiographer markers, and scanner stamps that correlate with disease prevalence or machine models. If high-performing vision transformers satisfy clinical benchmarks by covertly reading corner annotations rather than inspecting lung tissue, clinical deployment risks catastrophic diagnostic failure when transferred to unannotated scans. This study proves that output-level validation is insufficient for safety-critical medical ML and demonstrates the necessity of mechanistic circuit audits.

Caveats: The mechanistic investigation focused on text annotations in chest radiographs; other subtle imaging artifacts, such as rib-cage positioning or soft-tissue contrast differences across scanner vendors, may rely on distributed circuits that are harder to isolate to single attention heads. Additionally, developing automated, scalable intervention mechanisms to surgically excise these heads without degrading relevant visual representations remains an open challenge.