Wk2 - Advanced Contouring Flashcards
the different contour comparison metrics
- volume
- centre of volume
- volume overlap
- 2D shape and dimension
- 3D shape and dimension
contour comparison metrics - volume + limitations
- the absolute volume (cc)
- can be used to calculate the % volume different from the reference, mean volume and SD
limitation
- no spatial correlation of volumes
contour comparison metrics - centre of volume (COV) + limitation
- very easy to work out in TPS
- autoplace POI in ROI
limitations
- differing volumes can have the same COV
contour comparison metrics - volume overlap + limitation
- dice similarity coefficient
- concordance index (CI) = volume of intersection/volume of consensus x 100%
limitation
- do not measure distances between volume edges
contour comparison metrics - 2D shape and dimension + limitation
- measure maximum distance in a particular dimension
- can measure slice by slice
- can use BEV display with contours displayed
limitations
- can have same maximum in a particular dimension with different volumes and COVs
contour comparison metrics - 3D shape and dimension + limitation
- hausdorff distance is commonly used
- mean distance to agreement is also commonly used
- calculates differences between surfaces for multiple points around the volume
limitation
- maximum distance, average distance and 95% CI measurements only
- no directional information
- irregular surfaces can result in errors
- particularly for radial methods (eg. polar coordinates)
StructSure
- contour analysis software
- integrated into ProKnow
ProKnow
- contour and plan review software
- calculates the sensitive and sophisticated StructSure accuracy score as well as simple metrics (dice coefficient, total volume, matching/missing/extra volumes, etc)
- displays and analyses variability
MIM Maestro
- contouring, image registration, plan adaptation software
- calculates a wide range of metrics
Matlab
- calculates dice, Hausdorff distance, STAPLE
3D Slicer-Slicer RT
- calculates dice and Hausdorff distance
factors affecting manual contouring outcomes
- windowing
- image interpretation skills
- limitations due to image quality
deep neural networks (DNNs)
- a type of artificial neural network (ANN)
Describe the rationale for auto-contouring
With the advancement of RT techniques came the need for detailed contouring of datasets required to drive optimisation. Manual contouring is time consuming and is prone to intra and inter-observer error.
How does the threshold auto-contouring tool work?
Quick and effective contouring tool where upper and lower CT numbers are selected and only those will be outlined
What is the challenge with fully automatic segmentation methods?
To determine how much user interaction is necessary
When is it suitable to implement this into clinical practice
Contours need to be given to build the model - what is the gold standard?
Describe the body atlas based method
Reference image associated with atlas contours is matched to CT via deformable registration. Atlas contours are then morphed to match patients image.
Define AI
Models algorithms or computer programs to tackle certain tasks requiring human intelligence
Define machine learning
Numerical algorithms and models established to analyse data and derive or learn decision-making capabilities to achieve certain tasks
Define deep learning
Large scale hierarchal models with multi-layer architectures to automatically generate comprehensive representations and learn complicated inherent patterns of data
Explain 3 learning styles of DL
- Supervised - learns relationship between input data and labels (requires labelled data) good for diagnosis, image segmentation
- Unsupervised - learns patterns in input data (unlabelled) good for anomaly detection in QA
- Reinforcement - learns to perform actions in response to environment to maximise a reward function good for decision making in ART
List 4 different contouring tools
- Manual
- Grey scale interrogation
- Body atlas based
- Statistical shape modelling
2 types of greyscale interrogation - CT
1 - threshold
2 - model based segmentation
body atlas based contouring process
- reference image associated with atlas contouring is “matched” to the patients’s image (CT or CBCT) via a deformable registration algorithm
- the resulting deformation field is used to morph atlas contours to match the patient image
manual contouring is
time consuming and prone to intra and inter observer error
analysis of contouring methods and tools
- manual contours of an expert is used as the gold standard
- clinical expertise
- knowledge of spatial relationships, normal anatomy and patterns of disease progression
- multiple experts preferred (due to intra and inter observer error)
- consensus contours pooling the expertise of multiple clinicians
- within a discipline and across disciplines (radiologist, ROs and RTs)
- assessment of intra and inter observer variability required
threshold techniques - contouring CT
- most commonly applied to segment anatomy on individual 2D slices of the 3D data set in radiotherapy treatment planning systems
- 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 ROI is detected/tracked and the ROI is outlined
- successful auto-outlining is dependent on image resolution, significant contrast between corresponding structures and a continuous surface
threshold technique - PET
- can use count or SUV voxel data
- very contentious issue
emerging solutions
- artificial intelligence
- machine learning
- deep learning models
common types of DNNs
- convolution neural networks (CNNs)/ fully convolutional network (FCN)
- widely used to extract image features for classification
- U-nets commonly used for segmenting/contouring images
