MPANet: Cross-plane attention fusion for multi-planar brain tumour classification with auxiliary segmentation

Abstract

Deep learning models for brain tumour classification from MRI typically use a single viewing plane, whereas radiologists routinely examine axial, coronal, and sagittal views together. We introduce the Multi-Plane Attention Network (MPANet), a compact dual-task model that fuses features from all three planes in a single forward pass through a novel Cross-Plane Attention Fusion (CPAF) module and also produces a tumour localisation mask. Each plane is processed by its own MobileNetV3-Small encoder; CPAF lets each plane attend to the other two through query–key–value attention with squeeze-and-excitation plane weighting, and a Plane-Consistency Regularisation (PCR) loss encourages the planes to agree on their class predictions. CPAF is designed for subject-matched (Mode A) triplets, but BRISC provides no subject-level cross-plane identifiers; all results therefore use intra-class (Mode B) triplets whose planes may come from different patients, so CPAF learns class-conditioned cross-plane complementarity rather than per-patient anatomical correspondence. On the BRISC test set (540 samples, four classes), the 3.78M-parameter MPANet achieves a macro F1 of 0.9833 and a weighted F1 of 0.9796, within 1.24 percentage points of the best single-task classifier (EfficientNetB0, 0.9920), while producing tumour masks with an mIoU of 0.4853. Replacing CPAF with feature concatenation causes training to collapse ($F_1=0.14$), removing PCR reduces the combined score by 0.0278, and multi-plane fusion improves mIoU by 0.0537 over an axial-only baseline. An expected calibration error of 0.0217 indicates well-calibrated confidence on this test set but does not establish clinical applicability, which would require external validation.