Accompanying data for the paper "Automatic Segmentation of Porcine Lung Tensile Specimens for Mechanical Characterization"

Tensile testing is commonly used to characterize the mechanical behavior of soft materials and biological tissues. 
However, for highly deformable materials, the specimen cross-section changes significantly during the experiment, making stress estimation challenging. 
Current biomechanical approaches generally rely on indirect assumptions such as incompressibility or material homogeneity, which are not always valid for heterogeneous tissues.
This work presents a hybrid framework combining hand-crafted image-processing techniques and deep learning for tracking specimen geometry during uniaxial tensile tests performed on porcine lungs.
A top-view camera was used to continuously record specimen deformation in order to estimate sample width throughout the experiment. A U-Net convolutional neural network was trained on 110 annotated image–mask pairs from 15 tensile tests, with on-the-fly data augmentation.
The proposed pipeline achieved a mean Dice coefficient of 0.899 and a mean Intersection over Union (IoU) of 0.835, with inference times below 5 seconds per image. 
These results demonstrate the ability of the model to accurately detect specimen boundaries under challenging experimental conditions. 
Extending the method to lateral-view imaging could additionally enable thickness estimation and dynamic cross-sectional area reconstruction, paving the way toward real-time stress assessment in soft tissue mechanics.