Thickness-aware multitask U-Net for filament segmentation from cytoskeleton to extracellular matrix

The actin cytoskeleton is a key regulator of cellular mechanics, which stress fibers forming the main framework for force generation and for transmitting forces to the nucleus. Through this role, they influence adhesion, migration, division and cellular phenotype. If segmentation artificially thickens or breaks these filaments, downstream morphometrics become biased. We present a deep learning framework based on a multitask U-Net for the segmentation of actin stress fibers in adipose stem cells imaged by confocal microscopy. SFEX-derived masks were used as computational references to ensure reproducible thickness measurements (mean ± SD: 0.672 ± 0.04 µm) while manual measurements showed high inter-observer agreement (ICC = 0.82). A classical U-Net produced reasonable segmentation metrics (Dice = 0.620, IoU = 0.458) but systematically overestimated filament thickness, with per-image overestimation reaching up to 330%. In contrast, the proposed multitask U-Net jointly predicts segmentation masks and local thickness maps, reducing over-thickening to approximately 10–75% per image and closely reproducing the reference filament width distribution (Dice = 0.635, IoU = 0.469). Cross-domain tests on second harmonic generation collagen images further demonstrated structural generalization, with predicted fiber thickness differing by only 3.3% from binarized images and 14.5% from computational references, while maintaining fibrillar organization despite contrast differences. These results show that explicitly incorporating thickness prediction into a multitask architecture preserves fine cytoskeletal structures and extends to extracellular fibrillar networks such as collagen, enabling biologically faithful and reproducible segmentation that support quantitative analyses of cytoskeletal organization and cellular mechanics.