view article
Figure 7
An identical cross-sectional slice of a zebrafish bone scanned in the Zernike-nanoCT setup of P05, labeled using the (a) Sensor3D model trained for the ANATOMIX dataset with 70 images and 32 batch size and applied to P05 data versus (b) the same model re-trained on 20 ground-truth images of the P05 dataset. Note that these data have different size, noise level and contrast range as compared with the ANATOMIX data. The as-trained model produced a compromised, though not totally incorrect, classification, mainly mislabeling edges (indicated by the black arrows and green regions) of the bone class that are incorrectly classified as background values. Whereas Sensor3D training on 20 images from scratch required 67 epochs, the use of a pre-trained ANATOMIX model, retrained with 20 ground-truth images of a P05 data, improved classification of both bone and LCN yet required only 37 epochs (1 h).

Journal logoJOURNAL OF
SYNCHROTRON
RADIATION
ISSN: 1600-5775
Volume 31| Part 1| January 2024| Pages 136-149
https://doi.org/10.1107/S1600577523009852
Follow J. Synchrotron Rad.
Sign up for e-alerts
E-alerts
Follow J. Synchrotron Rad. on Twitter
Twitter
Follow us on facebook
Facebook
Sign up for RSS feeds
RSS