Abstract

Deep learning for brain tumour analysis holds significant promise for Sub-Saharan Africa, where specialist neuroradiologists are scarce and MRI infrastructure is limited. However, models trained on high-income country datasets are rarely evaluated for explainability on African MRI data, raising concerns about clinical trustworthiness in resource-constrained settings. We present a systematic evaluation of ex- plainable AI (XAI) across a domain shift from BraTS 2021 (GLI, source) to BraTS-Africa (SSA, target), using a matched 146-subject design. We evaluate three 2D classifiers (ResNet18, EfficientNet-B0, ViT-Small) and four 3D segmentation models (UNet, Attention UNet, TransUNet, Swin- UNETR) under zero-shot and domain-adapted conditions, with XAI as- sessed via Grad-CAM, RISE, and Grad-CAM++ using overlap, faithful- ness, and pointing game metrics. Transformer-based architectures con- sistently outperform convolutional models in cross-domain performance and explanation quality: ViT-Small achieves a pointing game accuracy of 0.655 on SSA despite comparable classification AUC to CNN models, and Swin-UNETR attains the best adaptation DSC (0.8411 ± 0.027) with the most coherent XAI profile, being the only model whose attri- butions are simultaneously faithful and anatomically aligned. Critically, high faithfulness does not imply trustworthy explanations: CNN mod- els produce confounded attributions that are genuinely sensitive to the highlighted regions, yet those regions are not the lesion. These findings argue for mandatory XAI evaluation in clinical AI pipelines for African healthcare settings.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_053.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_053_supp.pdf

Link to Open Review

Open Review Page: https://openreview.net/forum?id=D4aVpdbHvF

BibTex

@InProceedings{OlaAyo_Benchmarking_MICCAISAT2026,
        author = { Olabode, Ayotola Hannah AND Iorumbur, Aondona Moses AND Raymond, Confidence AND Onifade, Olufade Williams},
        title = { { Benchmarking Explainable AI for Brain Tumour Segmentation and Classification in African MRI } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17265},
        month = {pending},
        page = {pending}
}


back to top