Radiomics and Use of AI for Prognostication
What is it and how not to do it !
Images encode information
The relationship of the “numbers” in the matrix can be used to infer “patterns”
It is known that these features will often occur in “clusters” which tend to associate with specific tumor features.
Issue with feature extraction
Features are heavily dependent on the scan parameters, machine and the reconstruction algorithm used.
Currently available RTTQA recommendations for image quality are not sufficient for radiomics data quality assurance.
Issues with feature extraction..
In this study, investigators investigated the impact of slice thickness and pixel spacing on radiomic feature variability.
As shown in the image unless these parameters are harmonized, there is significant variability in a large number of radiomic features.
Feature selection
For handcrafted radiomics, more than 100 radiomic features can be extracted from images by common packages.
As a result the dimensionality of the data is >> the number of observations available.
Training models using this high dimensional data results in an overfitted model.
Hence feature selection is often needed.
Penalization and Shrinkage
These are techniques which are designed to prevent overfitting models.
They shrink the predictor effect estimates towards zero and eliminate variables based on that.
As shown in the image to the right > at small sample sizes do the opposite of what they are supposed to do.
Sample size requirements
AI : What is it
Simply put AI - developing machines to mimic human behavior and intelligence - currently we have “Narrow AI” - agents designed to solve a specific problem
Machine learning : Methods which learn from data using “automatic” analytic model building.
Deep learning: A subtype of machine learning which utilizes artificial neural networks to learn from data.
All DL is AI but all AI is not DL
What do we mean by deep learning ?
- Input data is processed in multiple “layers” which are connected to each other.
Each layer has ability to recognize / create “features” e.g. shapes, color.
The process learns which feature to place on which level on its own.
The layers transform the data into the “outcome” we are interested in.
Learn high level features from data
Feature extraction as done manually can be eliminated.
Deep learning for prognostication
True deep learning for prognostication are uncommon.
Most use a convolutional neural network where the model is optimized to predict the hazard (this is done through the use of a different loss function.
Performance better than radiomics in limited datasets.
Learning encoded information
17587 radiographs to classify fractures.
Models predicted scanner model, scanner brand and order priority better than predicting fracture
Best predictive ability when image data combined with patient clinical data and hospital process features.
In blackbox DL models these process variables may be unknowingly leveraged as variables for prediction which may lower reliability
Acknowledgements
This presentation is heavily influenced by the extensive work done on prognostic and predictive modelling at the PROGRESS framework website available at https://www.prognosisresearch.com
Conclusion
We invite you to join us at CHAVI and conduct meaningful and generalizable research using radiomics & AI. (https://chavi.ai)
