Learning Robust Medical Image Segmentation with Inductive Bias

Abstract

Despite the success of transformer-based and convolutional neural networks in 3D medical image segmentation, current architectures exhibit limited generalisation on small datasets and under distribution shifts, especially when high-quality examples are scarce for specific structures. We introduce IB-nnU-Nets, a family of U-Net variants augmented with inductively biased filters inspired by vertebrate visual processing. Starting from a 3D U-Net backbone, we insert two 3D residual components into the second encoder block that implement on- and off-centre-surround convolutions with fixed, pre-computed weights and act as complementary edge detectors. Across multiple organ and tumour segmentation tasks, we show that equipping state-of-the-art 3D U-Nets with an IB block improves accuracy and robustness, with the strongest gains in small-data and out-of-distribution settings. The framework and trained IB-nnU-Net models are publicly available.