Paper number: 136 A deep learning approach to segmentation of the - - PowerPoint PPT Presentation

paper number 136 a deep learning approach to segmentation
SMART_READER_LITE
LIVE PREVIEW

Paper number: 136 A deep learning approach to segmentation of the - - PowerPoint PPT Presentation

Paper number: 136 A deep learning approach to segmentation of the developing cortex in fetal brain MRI with minimal manual labeling AE Fetit , A Alansary, L Cordero-Grande, J Cupitt, AB Davidson, AD Edwards, JV Hajnal, E Hughes, K Kamnitsas, V


slide-1
SLIDE 1

A deep learning approach to segmentation of the developing cortex in fetal brain MRI with minimal manual labeling

AE Fetit, A Alansary, L Cordero-Grande, J Cupitt, AB Davidson, AD Edwards, JV Hajnal, E Hughes, K Kamnitsas, V Kyriakopoulou, A Makropoulos, PA Patkee, AN Price, MA Rutherford, D Rueckert

Paper number: 136

slide-2
SLIDE 2

Context – Developmental Brain Mapping

The Developing Human Connectome Project (DHCP) aims to make major scien;fic progress by crea;ng the first 4D connectome map of early life.

slide-3
SLIDE 3

Segmenta5on – Ul5mate Goal

Develop a 3D structural segmenta;on pipeline for fetal brain MRI to support connectomics research.

slide-4
SLIDE 4
  • Rapid changes in morphology over narrow ;me-scales.
  • Changes in white/grey-maKer intensi;es also take place.

Segmenta5on - Challenges

slide-5
SLIDE 5

Successful in other medical imaging applica;ons

  • However, main difficulty is in the need for large annotated

ground-truth.

  • Whilst large public datasets exist, they tend to mainly

include adult brain scans e.g. UK Biobank.

Deep Learning?

slide-6
SLIDE 6

Apply Draw-EM with fetal atlas to generate preliminary 3D labels Manual QC Train a multiclass 3D CNN using scans that passed QC step Apply the multiclass 3D CNN Give ~300 2D slices to expert annotator Refine the cortex labels Fine-tune the 3D cortex segmentation CNN

Minimal Labeling Workflow

Train a 3D cortex segmentation CNN

slide-7
SLIDE 7

Ka Kamnitsas e et al. 2016 t al. 2016

  • 3D

3D mo modeling deling u usin sing g DeepMedic DeepMedic

CNN Architecture

  • Th

Three parallel pathways:

  • no

normal rmal reso solu;o lu;on n

  • do

downsample wnsampled b by 3 y 3

  • do

downsample wnsampled b by 5 y 5

  • 8 la

8 layer ers per pa s per path thway y

  • Tr

Training batch size was set to 5

  • Learning r

Learning rate f e follo llowed a pr ed a pre-defined schedule. e-defined schedule.

slide-8
SLIDE 8

Preliminary mul5class labels

slide-9
SLIDE 9

Gestational age: 27.5 weeks

Example cortex refinement

slide-10
SLIDE 10

Example cortex segmenta5on

Gestational age: 28 weeks

slide-11
SLIDE 11

Thank you! Questions?

Paper number: 136