AI In RT Flashcards
Deep learning
large scale hierarchical models with multi-layer architectures to automatically generate comprehensive representations and to learn complicated inherent patterns of data
artificial neural networks, inspired by the brains network of neurons, various architectures
Includes: convolutional neural networks (CNN) and generative adversarial networks (GAN)
what is AI
Models, algorithms or computer programs designed by humans to tackle certain tasks requiring human intelligence can be generally considered as AI
It is a broad term covering ML/DL and other aspects of data science - often used interchanged
what is ML
subcategory of methods within the broad scope of AI
statistical methods of learning from data
ML approaches
supervised learning
unsupervised learning
reinforcement learning
transfer learning
ML processes
data can be collected from a variety of sources
data preparation
data wrangling or data cleaning
analysis through graphical plots and visualisations
explore basic relationships and patterns
Model selection
Model testing
Performance metrics
Validation and clinical commissioning
applications of ML in RT
RT simulation - brain pseudo CT for treatment planning
Contouring - auto segementation of target volumes and OAR in breast cancer patients
Dosimetry - curative intent radiation for prostate
QA - detecting MLC errors
Follow up - predictive modelling of survival and toxicity in patients following RT
Implications for clinical professionals
- reduction of repetitive low value tasks
- increased ability to make complex decisions using decision aids
- more personalised treatments
- maintenance of AI and data systems
-decision aids for RO and RTs - quality assurance applications most useful for RT and MPs
- auto contouring
- safe and effecient clinical practice
- patient outcomes should exceed
how can AI help clinically
deal with large amounts of data
find patterns and relationships previously too difficult for humans
develop decision aids for complex situations
remove menial repetitive tasks
reducing decision and process burden
how can ML help clinically
data input –> ? —? output
DVH data —> ML model —> Toxicity
model selection and training
to train supervised model to make predictions on new data - 70-80% original data subset required
learns patterns and begins to fit the model as close as possible to the data
supervised learning
-uses labelled data to help the model learn relationships present
- labelled data has both the input variables and the output variable of interest in the data set
- described as supervised because it requires human input to label the data
unsupervised learning
uses un-labelled data
- models are focused on finding patterns or groups present
- useful for dimensionality reduction, or reducing the number of variables
- called unsupervised as it doesnt require human labelling of outcome
semi-supervised learning
data contains some labels
allows the model to learn the relationship between the input and output variables but is not limited to looking at this connection
reduces the burden of labelling a large data set
reinforcement learning
- uses trial and error type approach through an agent that interacts with the environment
-The agent is set a task and is guided by reward and punishment as it makes decisions on how it
approaches the task
* Requires a large amount of domain knowledge to set up the model
* While similar to a human approach to perform a task, the model may address a problem in a way
that humans wouldn’t normally consider but could be beneficial
transfer learning
Uses a model trained for one task and applies it to a similar but different task
Reduces the need to re-train large & complex deep learning models
* Relies on the variables in the first task to be general and relatable to the second task
* Pre-trained models can be used in whole or part & adapted to a new setting
* Provides potential efficiencies in model development
* Can be both supervised and unsupervised
supervised learning methods
- classification: predicts a class outcome. e.g. yes OR no, Toxicity grade
- regression: predicts a numerical outcome. e.g. OAR dose volume
linear versus logistic regression
linear: fit a line, calculate R2 and adjust line, use formula to make prediction
logistic: fit a curve, calculate likelihood, adjust curve and calculate likelihood, repeat process multiple times, determine curve with maximum likelihood, make prediction using selected curve
Decision trees
ready to apply new data to predict outcome
easy to build, interpret and use
limited flexibility when used with new data
representation of possible solutions to a decision based on certain conditions
random forest
utilises decision trees
same process repeated 100s of times
run new data through each decision tree and then aggregate results
process of using bootstrapped data and aggregating the results = bagging
more flexible than decision trees and can adapt new data more easily
low interpretability - due to large number of trees used
can be used for classification and regression
neural networks
idea based on the human brain (neurons)
can involve multiple hidden layers
many different architectures available
combines many different mathematical functions to create complex lines of best fit
limitations
high bias
high variance
How do we ensure safety of patients is maintained
Relies on an understanding of how AI systems are generated and applied
Staff education on AI basics
Ability to understand assumptions and limitations of an AI application to ensure accountability in decision making
Communication with patients about how these tools are utilised
Hype cycle parts
- Technology trigger
- Peak of inflated expectations
- Trough of disillusionment
- Slope of enlightenment
- Plateau of productivity
