Source-linked AI summary
Spatial Feature-wise Linear Modulation (SpFiLM) for Contrast Agent-Aware Brain Parcellation
Pushpendra Singh, Joshua R. Astley, Roman Rodionov, John Duncan, Tom Vercauteren, Rachel Sparks
TL;DR
Brain parcellation tools are generally developed for T1w MRI, while T1ce appearance differences are spatially non-uniform. This paper introduces SpFiLM, which combines image-derived spatial modulation with contrast conditioning; it reports the best performance on both contrasts and a mean Dice of 84.1 against the U-Net baseline.
Problem
T1w-trained parcellation tools have systematically different morphometric estimates on paired T1w and T1ce scans, whose appearance differences are not uniform.
Source: §1 · p. 2Method
SpFiLM combines image-derived spatial bases with contrast-conditioned coefficients to spatially modulate feature scales and shifts in a unified network.
Source: §2 · p. 3, §2.2 · p. 4Results
84.1 mean Dice, a gain of 3.9% over the U-Net baseline and 0.7% over channel-wise FiLM, with SpFiLM best on the full dataset and each contrast.
Source: §4 · p. 6Takeaways & Limitations
A single jointly trained T1w-and-T1ce network was the best-performing model on both contrasts and converted pooled data into a gain on clinically acquired T1ce.
Source: §5 · p. 8The single-centre, single-scanner cohort and atlas-generated reference labels limit generalisability and bound the model and findings.
Source: §5 · pp. 8–9
Abstract
from arXiv · showhide
Most automated brain parcellation tools are developed and validated on T1-weighted (T1w) MRI. Yet, some clinical workflows for which parcellation is relevant only use contrast-enhanced T1w (T1ce) MRI, on which T1w-trained models are less accurate. We present a unified network that parcellates both pre- and post-contrast agent T1w MRI reliably, trained on a combination of the two with conditioning that spatially modulates its response differently for each. Feature-wise Linear Modulation (FiLM) is a known approach for input-based modulation in networks. It applies a per-channel scale and shift uniformly across the input. However, the appearance change between pre- and post-contrast varies locally across the brain, making FiLM suboptimal for our use case. In this work, we introduce Spatial FiLM (SpFiLM), a conditioning layer whose modulation varies spatially, assembling a voxel-wise scale and shift from image-derived spatial patterns. Using a cohort of 134 patients with paired T1w and T1ce MRI parcellated into 106 classes, the addition of SpFiLM layers in a UNet increased the mean Dice on the test set of 25 patients from 80.2% to 84.1%, a 4.9% relative improvement. Adding SpFiLM layers led to the best performance on both pre- and post-contrast MRI, even when controlling for network parameter counts.
1 Introduction
Brain parcellation tools are commonly designed for T1w MRI, while T1ce changes appearance locally and can produce different morphometric estimates. The paper introduces SpFiLM to condition parcellation on contrast with spatially varying modulation.
1 Introduction
- T1w-trained parcellation tools produce systematically different morphometric estimates on paired T1w and T1ce scans, although T1ce helps delineate vessels, tumour margins, and active lesions.T1ce enhancement occurs where the blood-brain barrier is absent or disrupted.Source: §1 · p. 2
- A unified contrast-aware network could avoid target-domain weight adaptation while using shared anatomy and contrast-specific differences from limited training data.Adaptation approaches require a separate post-training stage using target-domain data.Source: §1 · p. 2
- Global channel-wise FiLM treats image differences as uniform, whereas T1w-to-T1ce appearance differences vary locally across the brain.Domain randomisation may discard systematic, anatomically localised differences that conditioning could use.Source: §1 · p. 2
- SpFiLM uses spatially varying feature scales and shifts, can recover channel-wise FiLM as a special case, and consistently improves parcellation on both contrasts.The figure describes image-derived spatial bases recombined using coefficients generated from the contrast indicator.Source: §1 · p. 2, Figure 1 · p. 3
2 Methodology
The method supplies each scan and its known contrast indicator to a network that predicts voxel labels. SpFiLM combines image-derived spatial patterns with contrast-conditioned coefficients to modulate features within the brain.
2 Methodology
- Each unpaired scan is input with its known contrast indicator, and the network predicts a dense voxel-wise label map without receiving both contrasts jointly.The indicator comes from the acquisition record and is not predicted.Source: §2 · p. 3
2.1 Channel-wise FiLM
- Channel-wise FiLM applies one scale and shift per feature channel, with values generated from the contrast indicator and shared across every voxel in that channel.A three-layer MLP maps the indicator embedding to 2C scale and shift values.Source: §2.1 · p. 3
2.2 Spatial FiLM
- SpFiLM factorises spatially varying scales and shifts into image-derived spatial bases and contrast-conditioned coefficients, with K=0 recovering channel-wise FiLM.Separate basis generators produce scale and shift patterns; brain-mask gating confines modulation to relevant brain voxels.Source: §2.2 · pp. 4–5
2.3 Implementation Details
- The implementation uses a 3D U-Net and trains on combined, unpaired T1w and T1ce scans, with patient-level train, validation, and test splits.The dataset split is 97/12/25 patients, yielding 194/24/50 scans.Source: §2.3 · p. 5
3 Experimental Design
The study uses paired pre- and post-gadolinium scans from a single clinical cohort, with atlas-generated labels. It compares SpFiLM with controlled alternatives sharing its U-Net backbone.
3.1 Dataset
- The dataset comprises 134 patients with drug-resistant epilepsy scanned with and without gadolinium; patients with anatomy-distorting pathology were excluded.The scans used 1 mm isotropic MPRAGE acquisition.Source: §3.1 · p. 5
- Atlas-derived labels were generated on T1w scans and propagated to T1ce after rigid registration, leaving sub-voxel registration error as the dominant label-noise source.T1ce labels were propagated by nearest-neighbour interpolation because gadolinium enhancement precluded atlas fusion on T1ce.Source: §3.1 · pp. 5–6
3.2 Model Comparisons
- Controlled comparisons used the same U-Net backbone for unconditioned, parameter-matched, channel-wise FiLM, and attention baselines.Table 1 reports test-set Dice for 25 T1w and 25 T1ce scans, along with parameter counts and HD95.Source: §3.2 · p. 6, Table 1 · p. 6
4 Experimental Results
SpFiLM achieved the strongest reported parcellation performance across the full test set and each contrast, while ablations associate its gains with spatially varying contrast conditioning. Joint training especially benefited T1ce, and factorisation retained performance with fewer conditioning parameters.
4 Experimental Results
- 84.1 mean Dice made SpFiLM the best-performing model on the full dataset and on T1w and T1ce individually, 3.9% above U-Net and 0.7% above channel-wise FiLM.Differences from every baseline and state-of-the-art model were significant on both contrasts; HD95 differences were significant on T1w for selected comparisons but not on T1ce.Source: §4 · p. 6
4.1 Comparison with the State of the Art
- On a plain U-Net, SpFiLM exceeded nnU-Net ResEnc-L and SwinUNETR; adding SpFiLM to the nnU-Net encoder further improved mean Dice, especially on T1ce.This supports backbone-agnostic conditioning rather than gains arising only from a weak backbone.Source: §4.1 · p. 7
4.2 Ablation Study
- A learned contrast embedding did not improve over frozen one-hot encoding on either contrast, consistent with gains arising from spatial factorisation rather than encoding choice.Source: §4.2 · p. 7
- K=1 was too restrictive and slightly underperformed channel-wise FiLM, whereas K≥2 recovered the full gain; K=8 used 5.0M conditioning parameters versus 181M at full rank.Table 3 also reports that shallowest-level-only conditioning reached 83.7 versus 84.1 for the proposed configuration.Source: §4.2 · p. 7, Table 3 · p. 8
- Pooling T1w and T1ce training data improved SpFiLM on T1ce by +0.9%, versus +0.1% for U-Net, while T1w gains were +0.4% and +0.3%, respectively.The models used the same pooled data, so the larger T1ce improvement was attributed to contrast conditioning exploiting cross-contrast information.Source: §4.2 · p. 8, Table 2 · p. 8
5 Conclusion
SpFiLM uses spatially varying feature modulation conditioned on image contrast, with channel-wise FiLM as a special case. The study reports improved performance over established baselines, but its evidence is bounded by a single-centre, single-scanner cohort and atlas-generated labels.
5 Conclusion
- SpFiLM’s spatially varying, contrast-conditioned modulation recovered channel-wise FiLM as a special case, and ablations attributed improvement to this conditioning rather than parameter count or encoding choice.Source: §5 · p. 8
- SpFiLM surpassed nnU-Net and SwinUNETR, and a single jointly trained network performed best on both contrasts while converting pooled data into a T1ce gain unlike the unconditioned baseline.Source: §5 · p. 8
- Generalisability is limited by the single-centre, single-scanner cohort and by atlas-generated rather than manually annotated reference labels, whose accuracy bounds the model and findings.The study excluded patients with anatomy-distorting pathology; evaluating pathological cohorts and validating across institutions and scanners remain next steps.Source: §5 · pp. 8–9