safetycritiquebearishDeep models routinely misclassify out-of-distribution inputs with high confidence in medical diagnostic settingsComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02435v1