VUD-Net: a compact uncertainty-feedback convolutional framework for MRI-based brain metastasis segmentation

Abstract

Precise delineation of brain metastases on magnetic resonance imaging (MRI) is a key step in radiotherapy planning and follow-up assessment. The task remains difficult because metastatic foci vary markedly in size, occupy only a small fraction of the brain volume, and often show weak contrast at tumor-edema interfaces. To address these issues in a compact conference-version framework, we introduce VUD-Net, an uncertainty-feedback dynamic convolutional network for metastasis segmentation. Instead of applying a fixed set of convolutional filters to all cases, the model derives case-adaptive kernels from a variational latent distribution. A lightweight dynamic feature block further combines deformable sampling, receptive-field diversity, and boundary-sensitive filtering to handle irregular contours and small lesions. The uncertainty branch estimates feature-level disagreement from latent samples and returns this reliability cue to the encoder-decoder pathway, where it guides local refinement and suppresses noisy skip information. Experiments on a private contrastenhanced T1-weighted MRI cohort and the BraTS 2021 benchmark indicate that the proposed model reaches DSC values of 85.78% and 90.72%, respectively, while improving boundary consistency over representative baselines.