Conférence
Notice
Langue :
Anglais
Conditions d'utilisation
@ LE STUDIUM 2021
DOI : 10.60527/yhx8-2050
Citer cette ressource :
LESTUDIUM. (2021, 23 mars). Dr Natalie Reznikov - Application of deep learning for segmentation of 3D images in biomineralization , in Innate immunity in a biomineralized context: trade-offs or synergies?. [Vidéo]. Canal-U. https://doi.org/10.60527/yhx8-2050. (Consultée le 12 mai 2024)

Dr Natalie Reznikov - Application of deep learning for segmentation of 3D images in biomineralization

Réalisation : 23 mars 2021 - Mise en ligne : 26 avril 2021
  • document 1 document 2 document 3
  • niveau 1 niveau 2 niveau 3
Descriptif

Modern 3D imaging methods in biomineralization –such as X-ray tomography and dual-beam electron tomography – produce datasetsthat are rich in fine detail and enormous in size, often containing inevitableartifacts.  Rendering segmentations ofsuch datasets is a daunting task.  Therecent introduction of artificial neural network-based deep learning into bioimaginghas made 3D segmentation reliable, accurate and fast.  A highlight of convolutional neural networks(CNNs) is that artificial "neurons" are interlinked hierarchically, similarlyto how feature-forming patterns of an image are related.  Accordingly, when a raw image is presented toa deep net, the neurons of different layers perceive the patterns of differentcomplexity.  Upper-level neurons detect smallpatterns within their local context, and the local context itself forms patternsfor deeper neuronal layers, and within a larger context, and so on.  Thus, identification of features based on overt(e.g. contrast, gradient) and covert patterns(e.g. level of noise, waveletfrequency) becomes not only accurate, but also generalizable.  Once image patterns can be accurately enoughidentified as being features of interest – and thus the CNN is “trained” – suchpatterns can be segmented automatically on any similar image.  In machine learning, as in biologicallearning, the accuracy of pattern detection and classification improves withexperience.  Once trained, a CNN can betreated like an image filter – easy to preview, fast to apply, simple to share,and handy to reuse.  In thispresentation, I will explain the essence of deep learning and CNN operation fornon-computer scientists, and will illustrate this with examples of “difficult”3D images (a chick embryo inside a fertilized egg, and coral).

Dans la même collection

Sur le même thème