Paper Type
Complete
Abstract
High benchmark accuracy alone is no longer sufficient for clinical adoption of AI-assisted diagnostic systems. This paper presents a three-component trust evaluation framework covering explanation coherence, calibration quality, and governance operationalization. We apply this framework to a Vision Transformer (ViT) model fine-tuned on a brain tumor Magnetic Resonance Imaging (MRI) dataset consisting of 7,023 images, with a ResNet-18 baseline for comparison. The ViT achieves a test-set macro F1 of 0.989 and an AUC of 0.998. An attention faithfulness ratio of 168.2 indicates that rollout saliency maps identify the regions that materially influence model decisions. After temperature scaling, Expected Calibration Error (ECE) falls to 0.0166, and a governance policy at t = 0.90 auto-accepts 95.8% of cases with an error rate bounded at 0.159%. Conformal prediction achieves 92.1% empirical coverage. This framework offers Information Systems (IS) researchers a reusable protocol for transforming model outputs into actionable, deployment-ready governance specifications.
Paper Number
1839
Recommended Citation
Boit, Sorio; Mungania, Sharon; and Patil, Rajvardhan, "Toward Trustworthy AI in Clinical Imaging: Explainability, Calibration, and Governance Analysis of Vision Transformers for Brain Tumor MRI Classification" (2026). AMCIS 2026 Proceedings. 25.
https://aisel.aisnet.org/amcis2026/sig_dsa/sig_dsa/25
Toward Trustworthy AI in Clinical Imaging: Explainability, Calibration, and Governance Analysis of Vision Transformers for Brain Tumor MRI Classification
High benchmark accuracy alone is no longer sufficient for clinical adoption of AI-assisted diagnostic systems. This paper presents a three-component trust evaluation framework covering explanation coherence, calibration quality, and governance operationalization. We apply this framework to a Vision Transformer (ViT) model fine-tuned on a brain tumor Magnetic Resonance Imaging (MRI) dataset consisting of 7,023 images, with a ResNet-18 baseline for comparison. The ViT achieves a test-set macro F1 of 0.989 and an AUC of 0.998. An attention faithfulness ratio of 168.2 indicates that rollout saliency maps identify the regions that materially influence model decisions. After temperature scaling, Expected Calibration Error (ECE) falls to 0.0166, and a governance policy at t = 0.90 auto-accepts 95.8% of cases with an error rate bounded at 0.159%. Conformal prediction achieves 92.1% empirical coverage. This framework offers Information Systems (IS) researchers a reusable protocol for transforming model outputs into actionable, deployment-ready governance specifications.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
SIG DSA