AI Flashcards
Examples of AI in RT
- Auto segmentation in OAR contouring
- Predictive model of survival and toxicity in patients
Implications for clinical professionals
MRPBA statement
- Education in meeting the future needs of the workforce, education providers must ensure their curriculum evolves
- Need to be transparent to patients regarding the use of AI
- must understand ow AI systems are designed and applied in practice
- understand assumption and limitations of AI application to ensure accountability in decision making
- understand technology and data science including privacy and security
- use and deploy AI systems, maintain and overview them and direct their application in practice
- have control of AI supported systems to ensure patient safety
Model development components
Data, model, predictive uncertainty
Data collection points
Pre treatment: co-morbidities, history
Simulation: CT/MRI images, motion management, patient surveys
Dosimetery: contours, image fusion, quality metrics
Treatment: image matching shifts, motion management, IGRT images, toxicity reporting
Follow up: toxicity, survival
Data wrangling
collect data and put it into format that is useful and appropriate for ML modelling
Why data QA
Garbage in, garbage out
Need to ensure data quality
Avoid biases, insufficient quantity and inappropriate qualitt
Data sampling issues
Selection bias: not representative of target population
Class imbalance: underrepresntation of a class in the outcome variable
Insufficient quantity: low sample size
Prejudice bias: human biases such as racial, gender
Data measurement issues
Noise: medical images where artefacts may lead to reduce image quality
Uninformative or less informative: records do not convey all the necessary information to solve the problem at hand data
Exclusion bias: deleting valuable data
Measurement bias: systematically favouring a particular result when observing or measuring variable in data
Survey bias: substantial missing, incomplete and inconsistent responses to surveys, questionnaires or interviews used to collect data
Observer bias: favouring information that does not contradict the researcher’s desire
Data annotation issues
Recall bias - labeling similar types of data inconsistently
Annotation bias - inaccuracies induced by human errors when labeling
Lateral variability (inter observer variability): difference in practice between physicians or instituitions
Longitudinal variability (clinical drift): difference in practice over time
Missing data types
Missing completely at random (MCAR)
Missing at random (MAR)
Missing not at random (MNAR)
How to mitigate missing data
- Multiple imputations statisitical method for estimating the missing values
- Can just delete the rows
- Data augmentation
Distributed learning
- Each site/department develops the models, they send out models to all departments
- Gets rid of issues with legislation, patient privacy and ethics
Model QA importance
To check model interpretability and explainability
Model agnostic interpretation
- Global interpretation: provide info on what variables are most important
- Local interpretation: which variable from an individuals data was most important for the data output
Predictive uncertainty two components
Aleatoric and epistemically uncertainty
