Publications & results

ACCEPTED POSITION PAPER · EXPLAINS 2026

Localization Gains from Explanation Regularization Should Not Be Interpreted as Faithfulness Gains

Amrin Akter · Mauro Dalla Mura · Diana Mateus · Dawood Al Chanti

Accepted at the 3rd International Conference on Explainable AI for Neural and Symbolic Methods.

Results and figure ↓Presentation details →

Better lesion localisation does not establish greater faithfulness.

In this brain-tumour MRI case study, Contrastive Grad-CAM Consistency (CGC) yields higher mean localisation scores than a cross-entropy baseline. The perturbation-based faithfulness metrics disagree: insertion favours CGC, deletion favours the baseline, and IROF is unchanged at the reported precision.

Study: 3,064 MRI slices from 233 patients · three tumour classes · single-channel ResNet-50 · five-fold patient-grouped cross-validation · fixed training seed.

Overall results · mean ± standard deviation across five patient-grouped folds
Property	Metric	Baseline	CGC
Predictive performance	Accuracy (%) ↑	93.63 ± 1.91	94.87 ± 1.56
Localisation	Pointing Game (%) ↑	27.24 ± 3.76	36.02 ± 4.00
Localisation	Relevance Mass Accuracy (%) ↑	11.17 ± 1.60	18.01 ± 2.70
Faithfulness	Insertion AUC ↑	0.73 ± 0.06	0.78 ± 0.07
Faithfulness	Deletion AUC ↓	0.36 ± 0.07	0.41 ± 0.07
Faithfulness	Iterative Removal of Features ↑	0.57 ± 0.04	0.57 ± 0.07

↑ Higher is preferred; ↓ lower is preferred. Comparisons are descriptive: no statistical significance claim is made.

Grad-CAM examples for meningioma, glioma and pituitary MRI slices. Four columns compare the baseline and CGC on original and horizontally flipped inputs; green contours mark annotated tumours. — **Qualitative Grad-CAM comparison.** Original and horizontally flipped inputs for the baseline and CGC models; green contours mark tumour annotations. “Contrastive” denotes CGC. These selected examples do not quantify transformation consistency. Open full-size figure ↗

What this result supports

Report transformation consistency, localisation and model reliance as separate evaluation targets. Improvement in one does not establish improvement in the others.

Scope: one 2-D MRI dataset, one backbone, one attribution method and one training seed. Transformation consistency was not measured quantitatively. Faithfulness scores depend on the perturbation protocol; the dataset has documented non-tumour cues. This is a methodological case study, not clinical validation or a new XAI method.

Source: the authors’ accepted camera-ready manuscript, Tables 1–2 and qualitative comparison figure. Supported by ANR-25-CE45-1573 (EQUR-XSM); computations used GRICAD. A public paper link will be added when available.