FPN-Mamba: Multi-Scale MRI Tumor Classification

The clinical picture

A child comes in with headaches and unsteady balance. The scan shows a mass in the posterior fossa. The question the radiologist has to answer, quickly, is what kind of tumor it is.

Medulloblastoma is the most common malignant brain tumor in children, making up roughly 15 to 20 percent of pediatric central nervous system tumors. Outcomes depend heavily on the stage at which it is caught, so a missed diagnosis at screening delays treatment that could have been curative.

Histopathology is still the gold standard for confirming what a tumor is, but it is invasive and it is not available at the screening stage. So the first real decision gets made from the image. And that is genuinely hard: tumors of different types can look remarkably alike in texture, intensity, margin sharpness and location.

Why it's hard

Deep learning models are already good at this task. Several published approaches add attention mechanisms to CNN backbones and report strong numbers.

But they share a structural weakness. The attention almost always operates on a single, heavily downsampled feature map. That is a problem for exactly the cases that matter most, because small lesions and subtle boundaries are best represented at high resolution. Compressing the image before you attend to it can suppress the very evidence you need. Prior work identified small lesions, low contrast and atypical location as the main sources of false negatives, which is consistent with this.

The obvious fix is to run attention at finer scales. The obstacle is cost: self-attention scales quadratically with sequence length, which makes it prohibitive at high resolution.

The idea

FPN-Mamba keeps the whole image pyramid instead of collapsing it, so a tumor that is invisible at coarse scales is still visible at finer ones. Then it uses a state space model rather than self-attention to capture long-range context, because state space models run in linear time. That makes global reasoning affordable at every pyramid level, including the finest one.

In short: do not throw away resolution, and use a sequence model cheap enough that you do not have to.

How it works

An EfficientNet-B2 backbone extracts four feature levels (C2 to C5) at resolutions from 56x56 down to 7x7. A Feature Pyramid Network fuses them top-down through lateral 1x1 projections to 256 channels.

At each pyramid level, the standard 3x3 convolution is replaced by a LocalityMixing block. A depthwise convolution captures local 2D structure, a bidirectional Mamba scan captures long-range context, and a learned sigmoid gate weighs the two per position, so the model decides for itself where local detail matters and where global context does.

The four pyramid outputs are then serialized and passed through a cross-scale bidirectional Mamba block that models dependencies between levels. Each level is pooled with Generalized Mean Pooling using a learnable exponent, concatenated into a 1024-dimensional vector, recalibrated by Squeeze-and-Excitation channel attention, and classified by a fully connected head.

Everything is implemented in pure PyTorch with no proprietary CUDA extensions, so the results are reproducible on ordinary hardware.

Evaluation uses 761 MRI images (131 medulloblastoma against 630 non-medulloblastoma, an imbalance of roughly 1 to 4.8) under 5-fold stratified cross-validation. Class imbalance is handled with a weighted sampler, weighted binary cross-entropy and label smoothing. Thresholds are set per fold using Youden's J statistic, clipped to keep the operating point in a high-sensitivity regime, since in screening a missed tumor costs more than a false alarm.

What happened

FPN-Mamba reaches AUC 99.35 percent and F1 94.54 percent, with sensitivity 96.95 percent and specificity 98.25 percent. It beats all three baselines on every metric, including a fine-tuned ResNet-50 at AUC 97.91 percent.

The number that matters clinically is not the AUC. Across the pooled folds the model misses about 4 of 131 medulloblastoma cases, against 7 for ResNet-50. That is a 43 percent reduction in missed tumors, and missed tumors are the failure mode with real consequences.

The improvement holds up statistically. Bootstrap 95 percent confidence intervals do not overlap, and the Wilcoxon signed-rank test returns p equal to 0.0625, which is the minimum achievable value with five paired folds and requires the model to win on every single fold for every metric. It does.

The ablation tells a clear story about where the gain comes from. The Feature Pyramid Network is the dominant contributor, adding 4.31 points of AUC over the backbone alone. LocalityMixing and the cross-scale Mamba each push sensitivity higher but cost some specificity, and the GeM plus SE head restores the balance. Each component does something distinct.

On a separate three-class benchmark (meningioma, glioma, pituitary) the architecture reaches AUC 99.50 percent, so the design is not tuned to one task.