Generative Chained Label Propagation: A Semi-Supervised Approach for Enhanced 3D Segmentation of the Inferior Alveolar Nerve

Abstract

Manual voxel-level annotation of 3D medical images in dental radiology is a time-consuming process, particularly for intricate structures like the inferior alveolar nerve (IAN). As a result, many institutions rely on sparse annotations, sacrificing accuracy for efficiency, which limits the performance of deep learning models for dense segmentation. In this work, we address the problem of dense label scarcity as well as dense segmentation. We introduce Generative Chained Label Propagation (GCLP), a novel semi-supervised algorithm that generates and iteratively refines high-quality dense labels from sparse annotations utilizing PosPadUNet and our new architecture, ARP-UNet, both using a new training regime we propose. By training the dense segmentation model on a hybrid set comprising ground truth labels and the highest-quality dense labels generated by GCLP, our method achieves an IoU score of 0.715, surpassing the previous state-of-the-art of 0.642. Code and checkpoints are available on Github.