Advances In Contouring

Question 1

Motivation for auto-contouring

Answer 1

• 3D treatment planning - integration of large datasets
• iMRT and VMAT: complex tumour volumes and extensive OAR tolerances, detailed contouring required to drive datasets

Manual contouring: time consuming, prone to intar and inter-observer error

Question 2

Analysis of contouring methods and tools

Answer 2

Manual contours of an expert is used as gold standard - clinical expertise and reasoning

Consensus contours pooling the expertise of multiple clinicians

Question 3

Contour comparison metrics

Answer 3

Compare volumes
COV - centre of volume
Volume overlap - DICE, does not measure distance between volume edges
2D shape and dimension - can have maximum in a particular dimension with different volumes and COVs
3D shape and dimension - Hausdorff- irregular surfaces can result in errors

Question 4

Contour analysis software

Answer 4

StructSure (Standard imaging) - integrated into ProKnow
MIM Maestro
Matlab
3DSlicer-SlicerRT

Question 5

StructSure

Answer 5

Integrated into ProKnow

Question 6

ProKnow

Answer 6

Contour and plan review software
Calculates the sensitive and sophisticated StructSure accuracy score as well as simple metrics (dice coefficient, total volume)
Displays and analyses variability

Question 7

MIM

Answer 7

Contouring, image registration, plan adaption software
Calculates a wide range of metrics

Question 8

Matlab

Answer 8

Calculates Dice, Hausdorff distance, STAPLE

Question 9

3Dslicer - slicerRT

Answer 9

Wide community of contributions
Module that is installed separately

Question 10

Contouring tools

Answer 10

Manual
Image greyscale interrogation
Body atlas based methods
Statistical shape modelling

Question 11

Factors affecting manual contouring outcomes

Answer 11

Windowing
Image interpretation skills
Limitations due to image quality

Question 12

Grey-scale interrogation

Answer 12

CT: threshold techniques, model based segmentation
PET: threshold techniques

Question 13

Threshold techniques

Answer 13

Most commonly applied to segment anatomy on individual 2D slices of the 3D data set in radiotherapy TPS

Upper and lower limits for the CT numbers are selected, essentially applying thresholds for CT data to be included in the ROI

A start point is identified on the image proximal to the edge of the ROI to be outlined, the edge of ROI is detected/tracked and the ROI is outline

Question 14

Threshold techniques - CT

Answer 14

Depends on image resolution, significant contrast between corresponding structures and a continuous surface
Auto-outlining using threshold limits often requires manual editing
Outlining structures on all of the 2D slices can be very time consuming

Question 15

Threshold Techniques - PET

Answer 15

Can use count or SUV voxel data
Very contentious issue

Question 16

Body atlas based process

Answer 16

Two step process:
Reference image associated with atlas contours is ‘matched’ to the patient’s image (CT or CBCT) via a deformable reg algorithm
The resulting deformation field is used to morph atlas contours to match the patient image

Question 17

Body atlas based methods

Answer 17

Raystation multi-atlas based segmentation
Eclipse smart segmentation
MIM at last segment
Elekta ABAS
Velocity AI
Brain lab iPlan

Question 18

Statistical shape based methods

Answer 18

Pinnacle SPICE: fully automated hybrid approach which combines several deformable registration algorithms with model-based segmentation and probabilistic refinement

Question 19

Emerging solutions

Answer 19

AI - artificial intelligence
ML - machine learning
DLM - deep learning models

Question 20

Artificial intelligence

Answer 20

Models, algorithms or computer programs designed by humans to tackle certain tasks requiring human intelligence can be generally considered as AI

Question 21

Machine learning

Answer 21

Subcategory of methods within the broad scope of AI

Question 22

DML

Answer 22

Large scale hierarchical models with multi-layer architectures to automatically generate comprehensive representations and to learn complicated inherent patterns of the data

Question 23

Deep neural networks (DNNs) and common types

Answer 23

A type of artificial neural network (ANN)

Common types:
- convolution neural networks/fully convolution all network
- widely used to extract image features for classification
-U-nets commonly used for segmenting/contouring images

Question 24

DNN model training, accuracy and validation

Answer 24

Depend on the quality of the segmentations used to train the model
Yet to be determined:
- minimum number of patient datasets required to develop the models
- due to limited datasets suitable for developing models it is important to follow robust resampling methods to train, cross-validate and test models
- commissioning and QA required in clinical setting: use of contour comparison metrics

Question 25

Automated contouring for CBCT, MRI

Answer 25

To date most automated contouring solutions in RT have been for CT
Solutions required for MRI and CBCT
MRI: contouring for planning (MRI only simulation)
IGRT and plan adaption decision support, dose accumulation

CBCT: IGRT and plan adaption decision support, dose accumulation (LINACs with onboard kV imaging)

Question 26

Why not use fully automatic contouring

Answer 26

Difficult to determine what is the gold standard of contouring for the model
When is it suitable to integrate it to clinical practice
Determining how much user interaction is neccessary

Question 27

Model based segmentation - pinnacle

Answer 27

• form of grey-scale interrogation
• utilises models of organs to segment contours on individual 2D slices of the 3D data set in RT TPS
• Need to select upper and lower limit for CT numbers (threshold to CT data that is included in ROI generation)

Limitations
* dependant on image resolution, signficant contrast between abutting structures and a continuous surface.
* Often requires manual editing