Scopus Indexed Publications

Paper Details


Title
Do deep chest X-ray classifiers explain where they look? A multi-architecture faithfulness benchmark with false-positive attribution audit on VinDr-CXR

Author
SM Zain, ERAM MAHAMUD, Md Assaduzzaman,

Email

Abstract

Background

Hospital teams and external reviewers increasingly inspect saliency overlays during pre-deployment chest X-ray AI validation. A plausible heatmap can accompany an incorrect prediction, misleading validators who conflate visual highlighting with model correctness.

Objective

We propose and apply a structured pre-deployment attribution audit (null-controlled false-positive review and attentive-versus-blind false-negative categorization) on VinDr-CXR across three configurations (DenseNet-121, ConvNeXtV2-Tiny, Swin-B + LoRA) under Integrated Gradients. This study offers a structured pre-deployment warning protocol, not a deployable classifier and not a claim of clinical readiness, for detecting post-hoc saliency that appears plausible despite incorrect predictions.

Methods

VinDr-CXR is a public chest-radiograph dataset in which multiple radiologists independently annotated bounding boxes for 14 thoracic findings. On a 15,000-image VinDr-CXR training subset with 2-of-3 reader-consensus boxes, we generated 1528 Integrated Gradients maps and evaluated annotation overlap (precision-mass, recall top-50%, AUC-mIoU). Type B false positives were compared to an analytic null; false negatives classified as attentive or blind. Robustness checks included LIME, weight randomisation, and ablation.

Results

Across configurations, saliency–box overlap remained modest on correct predictions (maximum precision-mass 0.153). Type B wrong-class false positives concentrated IG mass above the analytic null (pooled Δ +9.8%, p = 7.4e-16), with overlays highlighting real abnormal regions despite an incorrect label. Attentive false negatives reached 51.4% of Swin-B + LoRA misses, including all Cardiomegaly cases. DenseNet-121 marginally failed the Adebayo weight-randomisation check (ρ = 0.512), warranting caution relative to ConvNeXtV2-Tiny (ρ = 0.263) and Swin-B + LoRA (ρ = −0.164).

Conclusions

Within this single public dataset using coarse consensus boxes and post-hoc saliency, saliency-box overlap should not be treated as evidence of diagnostic correctness during pre-deployment review. Because the top-predicted-class IG rule excluded three clinically important findings (Pneumothorax, Consolidation, Atelectasis) and because all analyses rest on one dataset without external or prospective clinical validation, results describe a pre-deployment *warning signal*, not deployment readiness. Null-controlled attribution audit may complement, not replace, conventional error analysis before clinical deployment.

Keywords

Journal or Conference Name
Intelligence-Based Medicine

Publication Year
2026

Indexing
scopus