L30: Machine Learning, Neural Networks, and Data Science Transforming imaging

Question 1

Data: What are the 2 starting points for a radiology AI?

Answer 1

• a readily available dataset – selected dataset automatically defines which use case the algorithm will be focusing on
• an idea for an application – have specific use case in mind and need to gather the required data

before gathering the right data for a radiology AI algorithm, define the use case in detail and determine what the exact output should be

• ie. should (a) detect/localize the tumour, (b) segment the tumour (2D/3D), (c) grade (or survival prediction) the tumour

Question 2

What are some of the pitfalls related to the choice of validation metric?

Answer 2

inappropriate choice of the problem category

• ie. object detection confused with semantic segmentation – DSC is strongly biased towards single objects and is therefore not appropriate for measuring the detection of multiple objects

poor metric selection

• ie. neglecting the small size of structure – single-pixel differences can hugely impact the metric scores, which is especially relevant given higher inter-rater variability and the non-deterministic nature of AI algorithms

poor metric application

• ie. inappropriate aggregation scheme – hierarchical data structure is often neglected when aggregating metric values, which is especially important for different numbers of cases per variable

Question 3

What is machine learning and what does it do?

Answer 3

• iterative process – requires constant revision of data input and assessment of output
• relies on ‘ground-truth’ data set or baseline to compare to – ie. need MRI to prove whether spinal cord lesion is present when developing AI-enhanced CT images
• uses modelling, which can be applied to many types of data – ie. language, cinematics, x-rays, CT datasets