HSV-Net: hybrid state-space and self-supervised vision fusion with progressive contour learning for CT pulmonary nodule segmentation.
- 1Dexing People's Hospital, Dexing, Jiangxi, China.
Abstract
Accurate voxel-level segmentation of pulmonary nodules on chest CT is needed for diameter and volume measurement, growth tracking, and quantitative assessment in lung cancer screening. The task remains difficult: nodules vary in attenuation, size, and contact with vessels or pleura, and clinically relevant morphology often lies in a thin, ambiguous margin. Existing systems commonly trade contour fidelity for speed-large prompt-driven foundation models can be accurate but slow, whereas compact CNN or SSM decoders are efficient yet tend to smooth spiculated and ground-glass borders. We propose HSV-Net (Hybrid State-space-Vision Network), a dual-stream 3D network that is prompt-free after nodule localization: an SSM3D stream provides whole-patch context with linear-complexity state-space blocks, and an OrthoLoRA stream adapts frozen DINOv2 features on axial, coronal, and sagittal slices. A Hybrid Gated Mixer (HGM) fuses the two streams before FPN decoding, and Progressive Contour Learning (PCL) supervises overlap, signed-distance-field geometry, and uncertainty-weighted contour refinement. On a 10-patient QIN Lung CT/QIN-LungCT-Seg development subset (n = 12 tumors; nested patient-level 5-fold CV), HSV-Net obtained the highest mean Dice (90.1% ± 1.0%; 95% CI [88.9, 91.3]) and Boundary F1 (86.3% ± 1.3%) among the compared methods on 128 × 128 × 64 patches, with a latency of 58 ms. A standard full-volume nnU-Net v2 baseline reached 88.6% ± 1.0% Dice on the same QIN folds, still below HSV-Net. On an independent LIDC-IDRI benchmark with expert volumetric consensus masks (186 patients/294 nodules; patient-level 5-fold CV within LIDC, not QIN→LIDC transfer), HSV-Net achieved 87.8% ± 1.1% Dice and 83.6% ± 1.3% Boundary F1, outperforming full-volume nnU-Net (87.1% ± 1.0% Dice). LUNA16 was retained only as a proxy size/localization check with diameter-matched spherical references, not as contour external validation.
Related
- Single-view echocardiography-based deep learning for pulmonary hypertension detection: a multi-level supervised and explainable approach
- A novel morphological assessment approach for abdominal aortic aneurysms: automated segmentation and measurement using a semi-supervised deep learning algorithm
- Segmentation of brain tumors using a morphological Laplacian-enhanced active contour model on MRI
- Automated quantitative analysis can support the identification of progressive pulmonary fibrosis in fibrotic interstitial lung disease.
- Self-supervised pretraining for dermoscopic lesion segmentation with limited labels: A controlled comparison of initialisation strategies
BibTeX
@article{cheng2026hsvnet,
title = {{HSV-Net: hybrid state-space and self-supervised vision fusion with progressive contour learning for CT pulmonary nodule segmentation.}},
author = {Liang Cheng and Xinsheng Wan},
journal = {Frontiers in oncology},
year = {2026},
month = oct,
doi = {10.3389/fonc.2026.1964406},
url = {https://europepmc.org/article/MED/42903127},
note = {CC BY},
}
RIS
TY - JOUR
TI - HSV-Net: hybrid state-space and self-supervised vision fusion with progressive contour learning for CT pulmonary nodule segmentation.
AU - Liang Cheng
AU - Xinsheng Wan
PY - 2026
DA - 2026/10/06
JO - Frontiers in oncology
DO - 10.3389/fonc.2026.1964406
UR - https://europepmc.org/article/MED/42903127
AB - Accurate voxel-level segmentation of pulmonary nodules on chest CT is needed for diameter and volume measurement, growth tracking, and quantitative assessment in lung cancer screening. The task remains difficult: nodules vary in attenuation, size, and contact with vessels or pleura, and clinically relevant morphology often lies in a thin, ambiguous margin. Existing systems commonly trade contour fidelity for speed-large prompt-driven foundation models can be accurate but slow, whereas compact CNN or SSM decoders are efficient yet tend to smooth spiculated and ground-glass borders. We propose HSV-Net (Hybrid State-space-Vision Network), a dual-stream 3D network that is prompt-free after nodule localization: an SSM3D stream provides whole-patch context with linear-complexity state-space blocks, and an OrthoLoRA stream adapts frozen DINOv2 features on axial, coronal, and sagittal slices. A Hybrid Gated Mixer (HGM) fuses the two streams before FPN decoding, and Progressive Contour Learning (PCL) supervises overlap, signed-distance-field geometry, and uncertainty-weighted contour refinement. On a 10-patient QIN Lung CT/QIN-LungCT-Seg development subset (n = 12 tumors; nested patient-level 5-fold CV), HSV-Net obtained the highest mean Dice (90.1% ± 1.0%; 95% CI [88.9, 91.3]) and Boundary F1 (86.3% ± 1.3%) among the compared methods on 128 × 128 × 64 patches, with a latency of 58 ms. A standard full-volume nnU-Net v2 baseline reached 88.6% ± 1.0% Dice on the same QIN folds, still below HSV-Net. On an independent LIDC-IDRI benchmark with expert volumetric consensus masks (186 patients/294 nodules; patient-level 5-fold CV within LIDC, not QIN→LIDC transfer), HSV-Net achieved 87.8% ± 1.1% Dice and 83.6% ± 1.3% Boundary F1, outperforming full-volume nnU-Net (87.1% ± 1.0% Dice). LUNA16 was retained only as a proxy size/localization check with diameter-matched spherical references, not as contour external validation.
SN - 2234-943X
ER -