Preserving instance continuity and length in segmentation through connectivity-aware loss computation

Abstract: In many biomedical segmentation tasks, the preservation of elongated structure continuity and length is more important than voxel-wise accuracy. We propose two novel loss functions, Negative Centerline Loss and Simplified Topology Loss, that, applied to Convolutional Neural Networks (CNNs), help preserve connectivity of output instances. Moreover, we discuss characteristics of experiment design, such as downscaling and spacing correction, that help obtain continuous segmentation masks. We evaluate our approach on a 3D light-sheet fluorescence microscopy dataset of axon initial segments (AIS), a task prone to discontinuity due to signal dropout. Compared to standard CNNs and existing topology-aware losses, our methods reduce the number of segmentation discontinuities per instance, particularly in regions with missing input signal, resulting in improved instance length calculation in downstream applications. Our findings demonstrate that structural priors embedded in the loss design can significantly enhance the reliability of segmentation for biological applications.