EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

DQIRMask: dual-quality and instance-relation mask learning for fluorescent neuronal cell instance segmentation

Yi Zhang, Xinyu Wu, Yichi Wang, Liwen Zhang, Wenxuan Zhang, Huaiwen Cao, Huaming Zhu, Yumeng Jiang, Haibo Shi

Physica Scripta · 06.11.2026 · 101 (44) · 446002

DOI: 10.1088/1402-4896/aeb737 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 02.11.2026. Verfasst von Yi Zhang und 8 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Key Program of the National Natural Science Foundation of China, Medical and Health Science and Technology Project of Shanghai Municipal Health Commission.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Fluorescent neuronal cell instance segmentation detects individual cells and delineates each cell at the pixel level for cell counting, soma morphology measurement, and spatial-organization analysis. Weak contrast, nonuniform fluorescence, substantial shape variation, and frequent contact or overlap can produce incomplete masks, boundary leakage, merged instances, and confidence scores that do not reflect mask quality. We propose Dual-Quality and instance-relation mask (DQIRMask), a CenterMask-based framework that combines Crowded-Instance Relation Reasoning, a Prototype-boundary decoder (PBD), and Dual-quality calibration. On the Yellow subset of Fluorescent Neuronal Cells v2, the full model improves average precision (AP) by 3.2 points over the strongest BlendMask configuration and by 5.1 points over the strongest CondInst configuration; the corresponding AP 75 gains are 6.5 and 10.1 points. Across six backbone ablations, the full configuration improves AP by 0.1–1.5 points and AP 75 by 0.1–3.0 points over the matching CenterMask baselines. The standalone effects are backbone dependent: the PBD reduces AP for four of six backbones. On the Green and Red subsets, AP 75 improves in all ten backbone-subset comparisons, although AP and F 1 do not improve in every case. These results support the full DQIRMask configuration for boundary-sensitive fluorescent-cell segmentation and show that component-level effects vary across backbones.

Förderung

  • Key Program of the National Natural Science Foundation of China
  • Medical and Health Science and Technology Project of Shanghai Municipal Health Commission

Zitieren

APA

Zhang, Y., Wu, X., Wang, Y., Zhang, L., Zhang, W., Cao, H., Zhu, H., Jiang, Y., & Shi, H. (2026). DQIRMask: dual-quality and instance-relation mask learning for fluorescent neuronal cell instance segmentation. Physica Scripta, 101(44), 446002. https://doi.org/10.1088/1402-4896/aeb737

Vancouver

Zhang Y, Wu X, Wang Y, Zhang L, Zhang W, Cao H, et al. DQIRMask: dual-quality and instance-relation mask learning for fluorescent neuronal cell instance segmentation. Physica Scripta. 2026;101(44):446002. doi:10.1088/1402-4896/aeb737

BibTeX
@article{Zhang2026,
  author = {Zhang, Yi and Wu, Xinyu and Wang, Yichi and Zhang, Liwen and Zhang, Wenxuan and Cao, Huaiwen and Zhu, Huaming and Jiang, Yumeng and Shi, Haibo},
  title = {DQIRMask: dual-quality and instance-relation mask learning for fluorescent neuronal cell instance segmentation},
  journal = {Physica Scripta},
  year = {2026},
  volume = {101},
  number = {44},
  pages = {446002},
  doi = {10.1088/1402-4896/aeb737}
}
Kostenlos starten →