Is It Width at the Activation, Not Parameter Count, That Lets Small-Data Medical Models Generalize?

Two findings from medical imaging seem to contradict each other. On small chest X-ray datasets, bigger networks memorize hospital tags and scanner quirks and generalize worse than small ones. On cardiac MRI, a wider U-Net generalizes better across scanners than slimmer versions. Both claims cannot be about the same thing. This proposal tests one possible resolution. "Capacity" bundles two quantities that act differently when data is scarce. The first is the number of trainable parameters, which gives the model room to memorize. The second is the number of channels a feature map has at the moment it passes through a nonlinearity such as ReLU (the function that zeroes negative values). A narrow layer at that point irreversibly throws away signal.

Image super-resolution already has a design that separates the two, called pre-activation channel expansion or "wide activation". It projects features into many more channels just before the nonlinearity, then projects them back, and uses factorized convolutions so the parameter count barely grows. We propose a controlled study on few-patient cardiac MRI segmentation. Activation width and parameter count will be varied independently across training-set sizes from 5 to 100 patients. The prediction is that, at a fixed parameter budget, wider activation improves cross-site accuracy on thin structures such as the myocardium. A planted-shortcut probe will check whether that same width also makes models latch onto spurious cues faster.

Core Claim

Holding parameters and training data fixed, we expect widening only the activation point to recover most of the cross-site benefit that Improving the generalizability of convolutional neural network-based segmentation on CMR images obtained by doubling every channel (UNet-64). That paper trained on 3,975 UK Biobank subjects and widened the whole backbone, so the parameter count grew roughly quadratically with width. We further expect parameter-matched wide activation to degrade more gracefully than global widening as the number of training patients falls. If either prediction fails, the clean outcome is that "width helps" in that paper was really about parameters or about data volume. That would also be worth knowing, because it would contradict the information-bottleneck story the super-resolution literature tells.

Grounding in Prior Work

The method comes from super-resolution. Wide Activation for Efficient and Accurate Image Super-Resolution (WDSR) showed that width at the ReLU matters more than width of the identity path. Its WDSR-B block reaches 6–9× expansion through a 1×1 projection followed by a 3×3 convolution, and it beats EDSR on DIV2K (34.536 vs 34.457 PSNR) with fewer parameters (0.60M vs 0.78M). It also found Batch Normalization harmful with small patches and replaced it with Weight Normalization. Feature Distillation Interaction Weighting Network for Lightweight Image Super-resolution and Image super-resolution reconstruction based on attention and wide-activated dense residual network extend the idea to texture preservation. All three use large natural-image training sets. None asks what happens when there are fewer than 100 training subjects, and none measures generalization across acquisition sites.

The nearest medical precedent is A Lightweight Self-Ensemble Feedback Recurrent Network for Fast MRI Reconstruction. It places wide activation modules inside a 44k-parameter recurrent reconstruction network and reports 35.65 dB PSNR at 15% sampling on cardiac data. However, it never ablates the expansion ratio against sample size, and it addresses reconstruction rather than segmentation. Blockwise Self-Supervised Learning at Scale expands to 2048 channels before a local loss to prevent collapse, but only at ImageNet scale.

Two papers motivate the study by exposing the tension. Unveiling COVID-19 from Chest X-ray with deep learning: a hurdles race with small data finds that ResNet-18 and an 8-layer CNN overfit less than ResNet-50 on the 447-image CORDA set, which is evidence against capacity. The CMR generalizability paper finds the opposite for width.

One paper constrains the design. Better NTK Conditioning: A Free Lunch from (ReLU) Nonlinear Activation in Wide Neural Networks shows that wide nonlinear layers push the gradients of similar inputs apart, which speeds up optimization. That same property could separate inputs along scanner artifacts as readily as along anatomy, so the study must test for shortcut amplification rather than assume width is benign.

Research Program

We build a single 2D U-Net family in which three quantities are set independently. The first is the nominal channel width $c$ (32 or 64). The second is the expansion ratio $r \in \{1, 2, 4, 6\}$ applied only inside each residual unit before its nonlinearity, using WDSR-B-style factorization. The third is the total parameter count $P$, matched by adjusting the low-rank inner dimension. This yields a parameter-matched grid. In it, UNet-32 with $r=4$ has roughly UNet-32 parameters but UNet-64-or-greater activation width, and UNet-64 with $r=1$ is the global-widening reference. Normalization is held fixed across all arms; Weight Normalization is the default following WDSR. A separate ablation compares Batch, Group and Weight Normalization so that any normalization effect is not mistaken for a width effect.

Training uses ACDC subsets of $N \in \{5, 10, 20, 50, 100\}$ patients, with five random patient draws per $N$. The data pipeline follows the CMR generalizability paper: 1.25 mm resampling, intensity normalization, and ±30° rotation with 0.7–1.4× scaling. Evaluation uses held-out ACDC patients. Cross-site evaluation uses BSCMR-AS if access can be obtained (599 subjects, 6 centers, 9 scanners). Failing that, the reverse direction serves the same purpose: train on small UK Biobank subsets and test on ACDC.

The analysis fits the cross-site Dice gap as a function of $(N, P, r)$. Our hypothesis is that the generalization gap depends mainly on $P/N$, while boundary error depends mainly on $r$. To test the mechanism directly, we record two diagnostics per layer. One is the fraction of channels that are zero across a validation batch, measured at each activation. The other is the effective rank of post-activation features. If wide activation helps for the reason the super-resolution papers give, gains should coincide with higher post-activation rank, and they should concentrate on myocardium and boundary metrics rather than the bulk of the left ventricle.

Baselines

The naive reference is a standard UNet-32 with $r=1$, trained on the same subsets with the same augmentation, step budget and learning-rate search. Every other arm is reported as a difference from it. The first strong competitor is globally widened UNet-64. It is the configuration the CMR generalizability paper found necessary for cross-scanner transfer, reporting Dice 0.90 for the left ventricle, 0.81 for the myocardium and 0.82 for the right ventricle on ACDC after training on 3,975 subjects. We retrain it on each small subset under an identical tuning budget. Wide activation counts as beating it if it matches or exceeds UNet-64's cross-site myocardium Dice at about half the parameters for $N \le 20$, with non-overlapping confidence intervals across the five patient draws. A shallow-network baseline, in the spirit of Conv8 and ResNet-18 from the COVID chest X-ray paper, is a four-level U-Net at $c=32$. It tests the simple advice to shrink the model.

The strongest existing few-shot method on ACDC is Retrieval-augmented Few-shot Medical Image Segmentation with Foundation Models. It retrieves similar annotated slices with DINOv2 and feeds them to SAM 2's memory, and it reports left-ventricle Dice 0.8472, against 0.66–0.69 for SAM 2 with point prompts. We run it with exactly the same annotated patients as its memory bank. We expect it to win at $N=5$ on left-ventricle Dice. Our claim does not require beating it. It is the reference that shows whether a small from-scratch model is clinically competitive at all, and it lets us compare myocardium and boundary performance, which that paper does not report. The key ablation compares expansion ratio $r$ against simply adding the same number of parameters elsewhere, as extra depth or a larger inner dimension without widening at the activation. This isolates activation width from parameter count.

Why not just…?

Why not just use a smaller network? The COVID chest X-ray study supports this for classification. However, the CMR evidence suggests narrow segmentation models lose cross-site robustness. The shallow baseline answers this directly. If it matches wide activation on cross-site myocardium Dice at every $N$, our claim fails.

Why not just widen everything, as UNet-64 does? This is the most serious threat. If UNet-64 matches parameter-matched wide activation even at $N=5$, then parameter count is not what hurts small-data training, and the factorization buys nothing beyond lower memory. The parameter-matched grid separates these possibilities, and we will report that outcome plainly if it occurs.

Why not just add stronger regularization or augmentation? Weight decay and AugMix-style augmentation can be applied to any architecture, and Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains found AugMix hard to beat in sparse-data settings. We therefore cross every architecture arm with a tuned weight-decay setting and with AugMix. The claim survives only if the width effect persists under the best regularizer.

Why not just use a foundation model? Retrieval with SAM 2 is strong, but it needs DINOv2 and SAM 2 at inference time. It also has no way to adapt features when the imaging modality shifts. The head-to-head comparison above will show where small wide-activation models stand relative to it.

Proposed Experiments and Decisive Checks

The main experiment is the $(N, P, r)$ grid on ACDC with cross-site testing. It reports Dice for left ventricle, right ventricle and myocardium, 95th-percentile Hausdorff distance, and the cross-site drop, averaged over five patient draws. Two checks are decisive.

The first is the parameter-matched contrast. Either UNet-32 with $r=4$ closes at least half of the gap between UNet-32 and UNet-64 on cross-site myocardium Dice for $N \le 20$, with gains tracking post-activation rank, or the bottleneck explanation is rejected.

The second is a planted-shortcut probe, which tests the concern raised by the NTK analysis. In ACDC training images we insert a faint synthetic intensity tag correlated with the presence of a basal slice, then remove the correlation at test time. Separately, on CORDA, we add a corner marker correlated with the label. For each $r$, we measure how quickly the model's reliance on the tag rises during training and how much accuracy falls when the correlation is removed. If wide activation picks up shortcuts faster, the practical recommendation becomes "wide activation plus anatomical cropping", which we test by repeating the probe with lung-field or heart-region crops.

A secondary replication applies the same expansion-ratio sweep to SEFRN-style reconstruction on fastMRI knee data with 5–20 training volumes. This checks whether any width effect transfers from segmentation to reconstruction.