Spatially-Informed Multi-Stage Active Learning for Segmentation of Nucleus in Soft X-ray Tomography

Accurate nucleus segmentation is critical, as the nucleus operates as the cell’s control center, responsible for holding instructions vital to cellular health. This makes nucleus segmentation a central focus in the study of disease-related abnormalities. Advanced 3D imaging techniques, such as soft X-ray tomography (SXT) offer quantitative, label-free, high-throughput imaging of whole cells, providing key insights into nucleus malformations. While SXT is widely employed for analyzing structural changes in various cell types, automatic nucleus segmentation remains challenging due to variations in nucleus patterns, influenced by factors like DNA packing, nuclear pores, and disease-induced morphological changes. To overcome these challenges, we developed ANSeg (Automated Nucleus Segmentation), a deep active learning-based 3D nucleus segmentation model. Starting with a minimal base model trained on only 5 volumetric data, we incrementally incorporated additional data and retrained the model through an active learning strategy. This approach led to significant performance improvements and achieved an average Dice similarity coefficient (DSC) of 91.8%. Ablation studies revealed that stages 3 (incorporation of dead cells) and 5 (other cell types with the highest uncertainty of segmentation) contributed minimally to further performance gains. This emphasized the critical role of stages 1, 2, and 4 for effective training, resulting in an average DSC of 92.47%. Additionally, a comparison between active learning and random selection strategies revealed a 2% performance advantage for the active learning approach when number of data is limited. Our approach demonstrates robust performance in 3D nucleus segmentation across diverse cell types, making it a powerful tool for analyzing nuclear malformations involved in disease mechanisms.