Huff Reece D, Houghton Frederick, Earl Conner C, Ghajar-Rahimi Elnaz, Dogra Ishan, Yu Denny, Harris-Adamson Carisa, Goergen Craig J, O'Connell Grace D
Department of Mechanical Engineering, University of California, Berkeley, Berkeley, CA, 94720, USA.
Weldon School of Biomedical Engineering, Purdue University, West Lafayette, IN, 47907, USA.
Sci Rep. 2024 Aug 8;14(1):18401. doi: 10.1038/s41598-024-68875-w.
Image-based deformation estimation is an important tool used in a variety of engineering problems, including crack propagation, fracture, and fatigue failure. These tools have been important in biomechanics research where measuring in vitro and in vivo tissue deformations are important for evaluating tissue health and disease progression. However, accurately measuring tissue deformation in vivo is particularly challenging due to limited image signal-to-noise ratio. Therefore, we created a novel deep-learning approach for measuring deformation from a sequence of images collected in vivo called StrainNet. Utilizing a training dataset that incorporates image artifacts, StrainNet was designed to maximize performance in challenging, in vivo settings. Artificially generated image sequences of human flexor tendons undergoing known deformations were used to compare benchmark StrainNet against two conventional image-based strain measurement techniques. StrainNet outperformed the traditional techniques by nearly 90%. High-frequency ultrasound imaging was then used to acquire images of the flexor tendons engaged during contraction. Only StrainNet was able to track tissue deformations under the in vivo test conditions. Findings revealed strong correlations between tendon deformation and applied forces, highlighting the potential for StrainNet to be a valuable tool for assessing rehabilitation strategies or disease progression. Additionally, by using real-world data to train our model, StrainNet was able to generalize and reveal important relationships between the effort exerted by the participant and tendon mechanics. Overall, StrainNet demonstrated the effectiveness of using deep learning for image-based strain analysis in vivo.
基于图像的变形估计是一种重要工具，用于各种工程问题，包括裂纹扩展、断裂和疲劳失效。这些工具在生物力学研究中很重要，在生物力学研究中，测量体外和体内组织变形对于评估组织健康和疾病进展至关重要。然而，由于图像信噪比有限，在体内准确测量组织变形尤其具有挑战性。因此，我们创建了一种新颖的深度学习方法，用于从体内收集的一系列图像中测量变形，称为应变网络（StrainNet）。利用包含图像伪影的训练数据集，应变网络被设计为在具有挑战性的体内环境中最大化性能。使用人工生成的经历已知变形的人类屈肌腱图像序列，将基准应变网络与两种传统的基于图像的应变测量技术进行比较。应变网络的性能比传统技术高出近90%。然后使用高频超声成像来获取收缩过程中参与的屈肌腱图像。只有应变网络能够在体内测试条件下跟踪组织变形。研究结果揭示了肌腱变形与施加力之间的强相关性，突出了应变网络成为评估康复策略或疾病进展的有价值工具的潜力。此外，通过使用实际数据训练我们的模型，应变网络能够进行推广，并揭示参与者施加的力与肌腱力学之间的重要关系。总体而言，应变网络证明了使用深度学习进行体内基于图像的应变分析的有效性。
