TAJ-Net: a two-stage clustered cell segmentation network with adaptive joint learning of spatial and spectral information DOI Creative Commons

Qing Zhang,

Xiaohui Zhou,

Chunyan Wu

et al.

Biomedical Optics Express, Journal Year: 2024, Volume and Issue: 15(8), P. 4584 - 4584

Published: July 4, 2024

Pulmonary adenocarcinoma is the primary cause of cancer-related death worldwide and pathological diagnosis “golden standard” based on regional distribution cells. Thus, cell segmentation a key step while it challenging due to following reasons: 1) It hard for pure semantic instance methods obtain high-quality result; 2) Since spatial appearances pulmonary cells are very similar which even confuse pathologists, annotation errors usually inevitable. Considering these challenges, we propose two-stage 3D adaptive joint training framework (TAJ-Net) segment-then-classify with extra spectral information as supplementary information. Firstly, leverage few-shot method limited data mask acquisition avoid disturbance cluttered backgrounds. Secondly, introduce an strategy remove noisy samples through two networks one 1D network type classification rather than segmentation. Subsequently, patch mapping map results original images results. In order verify effectiveness TAJ-Net, build hyperspectral datasets, i.e., (3,660 images) thyroid carcinoma (4623 40 bands. The first dataset will be released further research. Experiments show that TAJ-Net achieves much better performance in clustered segmentation, can regionally segment different kinds high overlap blurred edges, difficult task state-of-the-art methods. Compared 2D models, image-based model reports significant improvement up 11.5% terms Dice similarity coefficient dataset.

Language: Английский

TAJ-Net: a two-stage clustered cell segmentation network with adaptive joint learning of spatial and spectral information DOI Creative Commons

Qing Zhang,

Xiaohui Zhou,

Chunyan Wu

et al.

Biomedical Optics Express, Journal Year: 2024, Volume and Issue: 15(8), P. 4584 - 4584

Published: July 4, 2024

Pulmonary adenocarcinoma is the primary cause of cancer-related death worldwide and pathological diagnosis “golden standard” based on regional distribution cells. Thus, cell segmentation a key step while it challenging due to following reasons: 1) It hard for pure semantic instance methods obtain high-quality result; 2) Since spatial appearances pulmonary cells are very similar which even confuse pathologists, annotation errors usually inevitable. Considering these challenges, we propose two-stage 3D adaptive joint training framework (TAJ-Net) segment-then-classify with extra spectral information as supplementary information. Firstly, leverage few-shot method limited data mask acquisition avoid disturbance cluttered backgrounds. Secondly, introduce an strategy remove noisy samples through two networks one 1D network type classification rather than segmentation. Subsequently, patch mapping map results original images results. In order verify effectiveness TAJ-Net, build hyperspectral datasets, i.e., (3,660 images) thyroid carcinoma (4623 40 bands. The first dataset will be released further research. Experiments show that TAJ-Net achieves much better performance in clustered segmentation, can regionally segment different kinds high overlap blurred edges, difficult task state-of-the-art methods. Compared 2D models, image-based model reports significant improvement up 11.5% terms Dice similarity coefficient dataset.

Language: Английский

Citations

2