Vision-language models frequently flip GBM versus metastasis diagnoses under evidence-preserving slice reordering and label-order swaps, with up to 67.8% flip rates, so static accuracy overstates clinical reliability.
Artificial Intelligence Review , volume=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Evaluating the Diagnostic Robustness of Vision-Language Models Under Visual and Textual Perturbations
Vision-language models frequently flip GBM versus metastasis diagnoses under evidence-preserving slice reordering and label-order swaps, with up to 67.8% flip rates, so static accuracy overstates clinical reliability.