A Comparison of Convolutional, Transformer, and State-Space U-Shaped Pipelines for Neuronal Soma Instance Segmentation in Label-Free 3D Microscopy
Bioengineering · 14.11.2026 · 13 (11) · 1356
DOI: 10.3390/bioengineering13111356 ↗Kurz gefasst
Fachartikel in der Zeitschrift Bioengineering, online erschienen am 14.11.2026. Verfasst von Seda Bayat Toksöz und 3 weiteren Autor:innen. Die Arbeit nennt 50 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Scientific and Technological Research Council of Türkiye.
Abstract (Originalsprache)
Instance segmentation of neuronal somata in label-free three-dimensional (3D) microscopy is challenging because of low contrast, crowding, and species- and modality-dependent appearance. We compared seven U-shaped semantic-to-instance pipelines (3D U-Net, nnU-Net v2, SwinUNETR, 3D TransUNet, U-Mamba, Swin-UMamba-3D, and an exploratory MedMamba-UNet) on 240 public volumes of six rat, mouse, and human species–modality subsets, using the released partition. Models were trained for 25,000 steps with three seeds, post-processing was selected on source-domain validation volumes, and native-grid predictions were scored with Dice, object-level F1, panoptic quality (PQ), and error categories. In-domain, nnU-Net v2 (Dice 0.672 ± 0.001, mean ± SD over seeds) and Swin-UMamba-3D (0.640 ± 0.006) ranked highest, and the six other pipelines scored lower than nnU-Net v2 in Holm-corrected volume-level tests of Dice and F1@0.5 (Hodges–Lehmann Dice differences −0.032 to −0.092). Compared with target-matched in-domain results, Dice decreased by 0.037–0.067 after human-held-out species transfer and by 0.139–0.243 after oblique-to-DGC modality transfer, with all unadjusted confidence intervals excluding zero. PQ (0.253–0.384) remained well below Dice, and missed and spurious somata dominated the errors, so individual soma recovery should be evaluated alongside voxel overlap. Because test volumes informed protocol development and biological source identifiers are unavailable, these comparisons are exploratory.
Förderung
- Scientific and Technological Research Council of Türkiye
Zitieren
APA
Toksöz, S. B., Işık, G., Şahin, G., & Akin, E. (2026). A Comparison of Convolutional, Transformer, and State-Space U-Shaped Pipelines for Neuronal Soma Instance Segmentation in Label-Free 3D Microscopy. Bioengineering, 13(11), 1356. https://doi.org/10.3390/bioengineering13111356
Vancouver
Toksöz SB, Işık G, Şahin G, Akin E. A Comparison of Convolutional, Transformer, and State-Space U-Shaped Pipelines for Neuronal Soma Instance Segmentation in Label-Free 3D Microscopy. Bioengineering. 2026;13(11):1356. doi:10.3390/bioengineering13111356
BibTeX
@article{Toksz2026,
author = {Toksöz, Seda Bayat and Işık, Gültekin and Şahin, Gökhan and Akin, Erdal},
title = {A Comparison of Convolutional, Transformer, and State-Space U-Shaped Pipelines for Neuronal Soma Instance Segmentation in Label-Free 3D Microscopy},
journal = {Bioengineering},
year = {2026},
volume = {13},
number = {11},
pages = {1356},
doi = {10.3390/bioengineering13111356}
}
EUVIMED