Medical physics · 2026
Cross-scan fusion network: A registration-based framework for annotation-efficient 3D ultrasound segmentation in low back pain assessment.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Chen Pengyu, Zeng Zixue, Zhao Xiaoyan, Meng Xin, Sheng Zhiyu, Satapour Maryam, Cormack John M, Bean Allison C, Nussbaum Ryan, Maurer Maya, Landis-Walkenhorst Emily, Kim Kang, Wasan Ajay D, Pu Jiantao
Аннотация
BACKGROUND: Chronic low back pain (cLBP) profoundly impacts quality of life, yet its underlying mechanisms remain poorly understood. Three-dimensional (3D) ultrasound imaging offers valuable insights into cLBP but poses challenges of manual annotation due to large data volume and poor image quality.
OBJECTIVES: We aim to develop and validate a novel approach called Cross-Scan Fusion Network (CSFN) for segmenting anatomical tissue layers in 3D ultrasound images.
MATERIALS AND METHODS: We analyzed 3D B-mode ultrasound volumes of the lumbar region, with six tissue layers annotated: dermis, superficial fat, superficial fascial membrane, deep fat, deep fascial membrane, and muscle. The dataset included 69 labeled scans from 29 subjects and 30 unlabeled scans from 10 subjects. Labeled scans were split into training (n = 19), validation (n = 10), and independent test (n = 40) sets. For annotation efficiency analysis, 5, 10, 15, and 19 scans in the training set were separately treated as annotated, with the remainder considered unannotated. CSFN leverages a VoxelMorph-style elastic registration network trained with a novel Projected Hausdorff Distance Loss (PHDL) to accurately register anatomical tissue layers across 3D scans. This supports both sample-efficient learning (CSFN-SEL) and semi-supervised learning (CSFN-SSL). In the CSFN-SEL, controlled wrap ratios are applied to pairs of labeled scans to synthesize realistic image-mask pairs, while in the CSFN-SSL, labeled scans are registered onto unlabeled scans to generate high-quality synthesized scans for augmentation. Both real and synthetic data are then combined to train an nnU-Net segmentor, enabling robust segmentation with minimal manual annotations. CSFN was compared to fully-supervised nnU-Net with and without augmentation, and SimCLR (nnU-Net backbone). Model segmentation performance was evaluated using the Dice coefficient, while registration quality was assessed using Dice and Average Symmetric Surface Distance (ASSD). Results were reported as mean ± standard deviation(SD) and compared using paired two-tailed t-tests on class-wise subject averages. Significance was set at 0.05 and adjusted for multiple comparisons.
RESULTS: CSFN-SEL consistently outperformed the fully supervised nnU-Net baseline across varying numbers of labeled training samples, improving the mean Dice coefficient from 69.34% to 74.62% (+5.28%, p-value < 0.05/3). CSFN-SSL further improved performance, achieving 79.33% (±1.96%) and 81.14% (±1.43%) with 5 and 19 training samples, respectively. With only five labeled scans, CSFN-SSL delivered a 9.51% performance gain (effect size d = 0.671, p-value < 0.05/3) with minimal annotation.
CONCLUSION: CSFN improves segmentation accuracy and robustness even with very few labeled scans, providing an effective and practical solution for advancing cLBP imaging and analysis.
CLINICAL RELEVANCE SATEMENT: CSFN enables accurate, automated segmentation of anatomical tissue layers in 3D ultrasound scans with minimal manual annotations, potentially accelerating the clinical evaluation and management of cLBP.
OBJECTIVES: We aim to develop and validate a novel approach called Cross-Scan Fusion Network (CSFN) for segmenting anatomical tissue layers in 3D ultrasound images.
MATERIALS AND METHODS: We analyzed 3D B-mode ultrasound volumes of the lumbar region, with six tissue layers annotated: dermis, superficial fat, superficial fascial membrane, deep fat, deep fascial membrane, and muscle. The dataset included 69 labeled scans from 29 subjects and 30 unlabeled scans from 10 subjects. Labeled scans were split into training (n = 19), validation (n = 10), and independent test (n = 40) sets. For annotation efficiency analysis, 5, 10, 15, and 19 scans in the training set were separately treated as annotated, with the remainder considered unannotated. CSFN leverages a VoxelMorph-style elastic registration network trained with a novel Projected Hausdorff Distance Loss (PHDL) to accurately register anatomical tissue layers across 3D scans. This supports both sample-efficient learning (CSFN-SEL) and semi-supervised learning (CSFN-SSL). In the CSFN-SEL, controlled wrap ratios are applied to pairs of labeled scans to synthesize realistic image-mask pairs, while in the CSFN-SSL, labeled scans are registered onto unlabeled scans to generate high-quality synthesized scans for augmentation. Both real and synthetic data are then combined to train an nnU-Net segmentor, enabling robust segmentation with minimal manual annotations. CSFN was compared to fully-supervised nnU-Net with and without augmentation, and SimCLR (nnU-Net backbone). Model segmentation performance was evaluated using the Dice coefficient, while registration quality was assessed using Dice and Average Symmetric Surface Distance (ASSD). Results were reported as mean ± standard deviation(SD) and compared using paired two-tailed t-tests on class-wise subject averages. Significance was set at 0.05 and adjusted for multiple comparisons.
RESULTS: CSFN-SEL consistently outperformed the fully supervised nnU-Net baseline across varying numbers of labeled training samples, improving the mean Dice coefficient from 69.34% to 74.62% (+5.28%, p-value < 0.05/3). CSFN-SSL further improved performance, achieving 79.33% (±1.96%) and 81.14% (±1.43%) with 5 and 19 training samples, respectively. With only five labeled scans, CSFN-SSL delivered a 9.51% performance gain (effect size d = 0.671, p-value < 0.05/3) with minimal annotation.
CONCLUSION: CSFN improves segmentation accuracy and robustness even with very few labeled scans, providing an effective and practical solution for advancing cLBP imaging and analysis.
CLINICAL RELEVANCE SATEMENT: CSFN enables accurate, automated segmentation of anatomical tissue layers in 3D ultrasound scans with minimal manual annotations, potentially accelerating the clinical evaluation and management of cLBP.