Volumetric Latent Diffusion for Brain Tumour Segmentation with Missing Modalities

Abstract

Brain tumour segmentation from magnetic resonance imaging normally assumes that all four imaging sequences are available, but in clinical practice one or more are frequently missing. Existing methods for incomplete inputs learn modality-invariant representations or distil from a full-modality teacher, and return a single deterministic mask. Gen- erative formulations instead model a distribution over plausible masks, but the cost of sampling at full resolution has largely confined them to two dimensions. This paper proposes a three-dimensional conditional latent diffusion framework for whole-tumour segmentation from incom- plete multi-sequence MRI. Images and annotations are first compressed by two volumetric autoencoders, which makes diffusion over whole vol- umes affordable, and sequences are dropped at random during training so that a single model handles every combination of missing sequences. Because the model samples rather than predicts, repeated draws also yield a voxel-wise map of disagreement at no extra training cost. Evalu- ations on 2,601 studies from BraTS 2023 and 2024 show that the method reaches 0.955 Dice with all four sequences and 0.904 with FLAIR alone, and that it is the most accurate of the four methods compared in eight of the fifteen configurations. The disagreement between draws widens as sequences are withheld, giving a signal that accompanies the prediction rather than replacing it