Purpose
Imaging biobanks are becoming increasingly popular and active efforts from a number of different institutions in Germany,
the UK,
and US are making significant progress in collecting,
storing,
and organizing large sets of image data.
These large cohorts of data contain very diverse data from different sources,
with different resolutions,
and quality which makes large scale analysis challenging.
In this study,
we aim to develop large-scale tools for processing and extracting meaningful bone information and statistics from a heterogenous set of CT data collected for...
Methods and materials
The training and validation data are created by a trained radiologist that manually contoured all of the bone regions in 179 images from 60 different patients.
These segmentations were augmented by adding noise and deformations and were then used to train a multi-layered convolutional neural network (CNN) until both the training and validation had above 95%.
Results
Our tools link the PACS directly to a computing cluster and allow the analysis to be run on any CT dataset immediately and scales linearly to 133,000 patients per hour.
The results were very robust against noise,
contrast,
different body regions,
energy,
and dose settings.
Compared to state of the art threshold and morphology approaches,
it had 1000 fewer misclassified pixels at TPR of 99%.
Conclusion
The study presents the basis for conducting large scale studies to extract meaningful quantitative information from imaging biobanks.
The use of the latest techniques in machine learning and neural networks brought a substantial improvement over standard image processing methods but suffer from the reputation of being more difficult to understand and very black-box like in nature.
As a beginning of one possible approach for looking into the black box we investigate the different intermediate stages of the convolutional neural network.
The figure shows using a...
References
[Petersen2013] Petersen SE,
Matthews PM,
Bamberg F,
et al.
Imaging in population science: cardiovascular magnetic resonance in 100,000 participants of UK Biobank - rationale,
challenges and approaches.
J Cardiovasc Magn Reson.
2013;15(1):46.
doi:10.1186/1532-429X-15-46.
[Bamberg2015] Bamberg F,
Kauczor H-U,
Weckbach S,
et al.
Whole-Body MR Imaging in the German National Cohort: Rationale,
Design,
and Technical Background.
Radiology.
2015;277(1):206-220.
doi:10.1148/radiol.2015142272.
[Ronneberger2015] Ronneberger O,
Fischer P,
Brox T.
U-Net: Convolutional Networks for Biomedical Image Segmentation.
Medical Image Computing and Computer-Assisted Intervention (MICCAI),
Springer,
LNCS,
Vol.9351: 234--241,
2015,
available...