FPN-Mamba: Linear-Time Bidirectional State Space Feature Pyramids for Small-Lesion Brain MRI Classification

Small lesions are where screening models have the least margin for error: sub-5 mm brain metastases are among the most frequently missed, and their evidence lives at a resolution that single-scale attention discards. We adapt FPN-Mamba, which pairs a Feature Pyramid Network with Bidirectional Mamba state-space models running in linear time O(L): at each pyramid level a LocalityMixing block fuses depthwise convolution with a bidirectional Mamba scan via a learned gate, and a cross-scale Mamba, Generalized Mean Pooling, and Squeeze-and-Excitation head aggregate the result. Against a re-implemented InceptentionNet [3], the natural self-attention competitor, under a nested-calibration protocol in which checkpoint selection, early stopping, and thresholding never touch the reported fold, FPN-Mamba reaches AUC 0.982 vs. 0.897, recall 0.914 vs. 0.835, specificity 0.971 vs. 0.815, and recall on the smallest lesions 0.870 vs. 0.788 (5 seeds × 3 folds, all p < 0.02). Isolating the mixing mechanism inside one network, a bidirectional Mamba mixer scales far better than a matched self-attention mixer as resolution grows: at 512px it is 3.7× faster per training step at comparable parameter count, the direct payoff of O(L) over O(L2) at the resolutions where small lesions remain visible.