Speaker
Description
Segmentation of synchrotron microtomograms (SRµCTs) is very challenging, both for algorithmic solutions and for domain experts. To characterize biodegradable bone implants based on automatic segmentation, DESY and Hereon investigated the scaling of the 2D U-net for high-resolution volumes using three key model hyperparameters (i.e., model width, depth, and input size). To utilize the 3D information from the SRµCTs, the prediction is made from multiple viewing directions and then fused by a voting method. In the evaluation, we compare the results by intersection over union (IoU). In summary, combined scaling of the U-net (i.e., all three model parameters are optimized together) and multi-axis prediction fusing with soft voting yields the highest IoU for the least abundant class. The multi-axes prediction allows the computation of uncertainty estimates with very low additional computational cost. Overall, the time needed to segment a single 3D SRµCT is reduced by an order of magnitude.
| I want to give an oral presentation. | yes |
|---|---|
| I want to present a poster. | yes |
