Scaling the U-net: segmentation of biodegradable bone implants in high-resolution synchrotron radiation microtomograms

Highly accurate segmentation of large 3D volumes is a demanding task. Challenging applications like the segmentation of synchrotron radiation microtomograms (SR mu CT) at high-resolution, which suffer from low contrast, high spatial variability and measurement artifacts, readily exceed the capacities of conventional segmentation methods, including the manual segmentation by human experts. The quantitative characterization of the osseointegration and spatio-temporal biodegradation process of bone implants requires reliable, and very precise segmentation. We investigated the scaling of 2D U-net for high resolution grayscale volumes by three crucial model hyper-parameters (i.e., the model width, depth, and input size). To leverage the 3D information of high-resolution SR mu CT, common three axes prediction fusing is extended, investigating the effect of adding more than three axes prediction. In a systematic evaluation we compare the performance of scaling the U-net by intersection over union (IoU) and quantitative measurements of osseointegration and degradation parameters. Overall, we observe that a compound scaling of the U-net and multi-axes prediction fusing with soft voting yields the highest IoU for the class degradation layer. Finally, the quantitative analysis showed that the parameters calculated with model segmentation deviated less from the high quality results than those obtained by a semi-automatic segmentation method.

Automated summary

Highly accurate segmentation of large 3D volumes, especially for challenging applications like the segmentation of synchrotron radiation microtomograms (SR mu CT) at high-resolution, requires reliable methods such as the scaling of U-net with model hyper-parameters. Through systematic evaluation, it is observed that a compound scaling of the U-net with multi-axes prediction fusing yields the highest IoU for the degradation layer. Ultimately, quantitative analysis shows that the parameters calculated with model segmentation deviated less from high quality results compared to semi-automatic segmentation methods.