Deep learning to facilitate bone segmentation using nanotomography imaging
What is it about?
This work shows how Zernike phase-contrast enhanced X-ray nanotomography (nanoCT) can reveal micro- and nano-porosity in bone. However, the obtained data are often plagued with halo and shade-off artifacts, which can make segmentation challenging. To address this issue, the authors proposed incremental deep learning (DL) training to classify bone porosity on reconstructed data. This method was found to be reliable and predictable, whereas the traditional threshold method failed to separate the voids and the outer edges of the bone. This study compares the outcomes of model training and validation with different training data and batch size parameters and suggests that the Sensor3D models are slightly more accurate than U-Net models within the limits of the training data and batch sizes tested.
Why is it important?
The major advantage of the Zernike-nanoCT method over other imaging methods is that bone can be imaged rapidly in three dimensions within a relatively large field of view, which makes it possible to repeat measurements on overlapping regions of interest, enabling even larger samples to be mapped rapidly in 3D. Combining Zernike-nanoCT with available DL libraries facilitates bone analysis, reducing human-based and time-consuming classification effort.