Most brain-parcellation tools expect a plain T1-weighted stack. Epilepsy, oncology, and stereotactic workflows often only hang contrast-enhanced T1 (T1ce) after gadolinium. SpFiLM is a single 3D U-Net that takes either series plus a contrast flag and returns a 106-region brain overlay without a separate T1ce adaptation stage. Pushpendra Singh, Joshua R. Astley, Roman Rodionov, John Duncan, Tom Vercauteren, and Rachel Sparks (King’s College London / UCL Queen Square) posted arXiv:2609.07718 around 7 September 2026. The paper is Spatial Feature-wise Linear Modulation (SpFiLM) for Contrast Agent-Aware Brain Parcellation. Code is at github.com/p-singh-kcl/spatial_film_parcellation.
Viewer shops already know the hang: T1 morphometry tools drift when the only series on the worklist is T1ce, because vessels, choroid plexus, and barrier-disrupted tissue light up locally. Retraining a second model, or forcing a domain-adaptation pass, is the usual workaround. SpFiLM keeps one network and tells it which contrast it is looking at.
Where the contrast change actually lives
Channel-wise FiLM scales and shifts each feature channel with one number per channel from the contrast indicator. That is a global dial. Pre- to post-gadolinium appearance is not global. It is local. SpFiLM keeps the contrast indicator, but builds voxel-wise scale and shift from image-derived spatial bases. Two small CNNs read a resampled copy of the scan and emit K spatial basis maps for scale and for shift. A tiny MLP turns the contrast flag into channel offsets plus mixing weights that recombine those bases. With K set to zero the layer collapses back to ordinary FiLM.
Modulation is gated by a brain mask from HD-BET so the spatially-invariant terms do not rewrite extracranial voxels. The backbone is a five-level 3D U-Net (widths 32 to 512) with deep supervision. Inputs are skull-stripped and z-normalised inside the mask. Training treats T1w and T1ce as unpaired: a patient’s two contrasts never sit in the same iteration. Inference uses a 128-cubed sliding window with 50% overlap and Gaussian patch weighting.
Where it was measured
The cohort is 134 adult drug-resistant epilepsy patients from NHNN London with paired 1 mm isotropic MPRAGE T1w and T1ce (REC 20/LO/0966). Cases with distorting lesions, tumours, or prior resection were excluded. Pseudo labels for 106 Mindboggle-style classes came from multi-atlas GIF fusion on T1w, then rigid nearest-neighbour propagation onto T1ce. The patient-level split is 97 / 12 / 25 train / val / test (194 / 24 / 50 scans).
On the held-out 25 patients, the plain U-Net averages 80.2% Dice across both contrasts. SpFiLM (K=8, 27.7M params) reaches 84.1% overall, with 86.6% on T1w and 81.6% on T1ce. Channel-wise FiLM lands at 83.4% overall. A parameter-matched wider U-Net barely moves the needle (80.4%), so the gain is not from parameter count alone. Dropping SpFiLM onto nnU-Net ResEnc-L pushes the combined score to 84.5%. HD95 on T1w falls to about 2.1 mm for SpFiLM and 1.93 mm for nnU-Net plus SpFiLM. Metrics stay in the body as support; the practical claim is one network for both hangs.
How this lands in a viewer
If you already hang brain MRI in a DICOM viewer, treat SpFiLM as a contrast-aware overlay path, not a new worklist. Read the contrast agent flag from the acquisition record (or from a tech checkbox when DICOM metadata is messy), run skull-strip, and hang the 106-region map as an editable overlay for volumetry, radiotherapy targeting, or SEEG planning. Fail closed when the series is not T1-family, when pathology has already warped anatomy outside the training exclusions, when the contrast flag is wrong, or when you need FreeSurfer-grade cortical thickness rather than GIF-style region labels. Labels on T1ce inherit rigid registration noise from the T1w atlas path, so treat boundary edits on enhanced vessels and choroid as expected cleanup, not a surprise.
GazeRefine turns fixations into overlays without retraining. This paper is pre- or post-gadolinium T1 in, shared brain-region overlay out with spatial contrast conditioning.
Rebuild from arXiv:2609.07718. As of 9 September 2026 the abstract and PDF respond (HTTP 200). The preprint is the source of record until a camera-ready version exists. Public code is linked in the paper.
Sources
- Pushpendra Singh, Joshua R. Astley, Roman Rodionov, John Duncan, Tom Vercauteren, Rachel Sparks. Spatial Feature-wise Linear Modulation (SpFiLM) for Contrast Agent-Aware Brain Parcellation. arXiv:2609.07718, posted ~7 September 2026. https://arxiv.org/abs/2609.07718 (HTTP 200 on 9 September 2026). PDF: https://arxiv.org/pdf/2609.07718 (HTTP 200 on 9 September 2026). Code: https://github.com/p-singh-kcl/spatial_film_parcellation.