What attention gates actually buy you
Standard U-Net against an attention gated one on BraTS 2021, everything else held fixed. The Dice scores come out identical. The boundary error does not.
10 mm
HD95, whole tumor
from 12 mm
5.5 mm
HD95, enhancing
from 7.5 mm
0.88
Dice, whole tumor
identical for both models
98%
Background voxels
the imbalance being fought
The question
Attention gates are widely said to help segmentation by suppressing irrelevant background activations. That is easy to accept and harder to test, because published comparisons usually differ in data, augmentation and schedule at the same time as the architecture. On BraTS 2021 the difficulty is concentrated in one place: roughly 98 percent of voxels are background, so a model can look strong on aggregate overlap while getting the boundary wrong.
Approach
One pipeline trains either architecture from the same configuration file, over the same data loading, augmentation, schedule and logging, so the architecture is the only thing that differs between runs. Four modalities stacked into one tensor, volumes resized to 128 by 128 by 144, DiceFocalLoss with inverse frequency and dynamic per batch class weighting, hyperparameters tuned with Optuna, five fold cross validation at 52 epochs per fold on two H100 GPUs, about ten hours a fold. Scored on Dice, sensitivity, Jaccard, HD95, precision and specificity rather than Dice alone. Joint project: I built the pipeline, the Attention U-Net and the visualisations, and my partner did the preprocessing, the Standard U-Net and the metrics.
Outcome
The two models are indistinguishable on overlap and clearly separated on boundary. Dice is identical to two decimals on all three regions, 0.82 enhancing, 0.86 core, 0.88 whole tumor, and so is Jaccard. HD95 is where the attention gates show up: 7.5 mm to 5.5 mm on enhancing tumor, 9 to 7 on core, 12 to 10 on whole tumor. Sensitivity improves slightly across the board and precision on tumor core goes 0.87 to 0.89, with a small specificity cost on whole tumor, 0.9980 to 0.9975.
What it taught me
If the report had carried Dice alone, the honest conclusion would have been that attention gates do nothing here, and that conclusion would have been wrong. The gates buy boundary fidelity, which overlap is close to blind to, and boundary is the part a surgeon or a radiotherapy plan actually acts on. It is the same lesson my registration work keeps returning: pick the metric that can see the effect you are claiming, then report the ones that cannot alongside it.