UM--ProtoShare: UNet-Guided, Multi-scale Shared Prototypes for Interpretable Brain Tumour Classification Using Multi-sequence 3D MRI

Ali Golbaf, Vivek Singh, Swen Gaudl, Emmanuel Ifeachor

Keywords: Brain Tumour Classification, Multi-sequence 3D MRI, Interpretable Deep Learning, Case-based Models

Abstract:

Deep learning shows strong promise in brain tumour classification using Magnetic Resonance Imaging (MRI), although limited interpretability constrains clinical translation.
Most interpretability methods are post-hoc and yield visual attribution maps that are only weakly connected to the decision process.
Clinicians prefer decisions built from evidence they can recognise and verify on MRI, rather than post-hoc explanations.
Case-based models embed reasoning by comparing image evidence with learned prototypes, yielding ``this looks like that'' rationales at decision time and mirroring clinical reasoning.
Building on this paradigm, we introduce UM--ProtoShare, which compares the input multi-sequence 3D brain MRI with a bank of shared, class-agnostic, multi-scale prototypes for pre-operative glioma grading.
It returns not only a label, but a set of prototype matches that highlight where the model found support for its prediction.
UM--ProtoShare uses a 3D ResNet-152 encoder, a lightweight UNet--style decoder with gated encoder--decoder fusions, and a normalised soft-masked mapping module to align and highlight prototype evidence on MRI.
On BraTS-2020, ablations show additive benefits from the normalised mapping module, prototype sharing, multi-scale prototypes, and the decoder with gated fusions.
Varying the allocation of prototypes across scales identifies a balanced accuracy--interpretability configuration that closely approaches a strong 3D ResNet-152 in classification performance (Balanced Accuracy: 88.40 $\pm$ 2.80; 1.48 percentage points lower) while delivering more faithful and spatially precise evidence than prior case-based models, with Activation Precision (AP) 88.72 $\pm$ 1.60 ($+$11.0\% vs MProtoNet; $+$4.0\% vs MAProtoNet) and Incremental Deletion Score (IDS) 5.10 $\pm$ 1.30 (lower is better, $-$32.3\% vs MProtoNet, $-$25.3\% vs MAProtoNet).

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Reproducibility: https://github.com/AliGolbaf/um-protoshare

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 52