Artificial intelligence to enhance personalized decision making

Question 1

What is artificial intelligence?

Answer 1

• “Artificial intelligence leverages computers and machines to mimic the problem-solving
  and decision-making capabilities of the human mind”.
• “Artificial intelligence is the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings” (Britannica)
• “Artificial intelligence makes it possible for machines to learn from experience, adjust
  to new inputs and perform human-like tasks” (SAS)

Question 2

Describe the definition of:
- artificial intelligence
- machine learning
- deep learning

Answer 2

• artificial intelligence → programs where human intelligence is incorporated into machines through an algorithm.
• machine learning → AI algorithm allowing systems to learn from data
• deep learning → subset of machine learning in which artificial neural networks adapt and learn from vast amounts of data, thus mimicing human brain-like behavior

Question 3

Skipped slide 8 -13

Question 4

What is generalization in regard to AI?

Answer 4

That the AI’s algorithm performs well on new, previously unseen data.

Question 5

What is a training error?

Answer 5

Error computed on the training set (during training we aim to reduce the training error)

Question 6

What is a generalization error? And how can this type of error be estimated?

Answer 6

• Generalization error → the expected error on the new, previously unseen data (goal is to minime the generalization error).
• By measuring the performance on a test data set which must be independent from the training set.

Question 7

During training of an AI algorithm, the training data set is split into multiple sets:
- training set
- validation set
- test set

For what are these types of sets used for?

Answer 7

• training set → used to optimize parameters/weights
• validation set → used to tweak hyperparameters such as learning rate
• test set → used for final evaluation as a surrogate of new and unseen data

Question 8

Name two reasons why the training data set needs to be large enough.

Answer 8

• To prevent overfitting to the training data
• So it generalizes well to new, unseen data.

Question 9

Typical medical images such as CT contain about 262.144 voxels on the 2D plane and 100.000.000 voxels with a 3D volume. The simplest fully connected classification network would have > 100.000.000 weights.
How is this problem tackled?

Answer 9

By applying image filtering

Question 10

What is image filtering?

Answer 10

• Transformation of the original image into a filtered image by applying some filtering operation → convolution.
• A filter slides over the image and performs a filtering operation at each pixel/voxel.

Question 11

What are convolutional neural networks?

Answer 11

Simply said: an image consists of millions of pixels. With the use of image filtering and convolution, the image with millions of pixels is turned into a filtered image (i.e. feature extraction). And with the use of deep learning, individual components (i.e. features) of the filtered image are classified/categorized (i.e. classification).

Question 12

An example is given where it is explained that with the use of CT imaging you can detect calcium deposits in coronary artieres (prediction of calcification/artery narrowing). However, CT imaging uses ionizing radiation and therefore is a potential health hazard.
A solution for this is to use low-dose CT imaging. However, this increases the amount of noise artifacts and thus reduces the accuracy of calcium scoring. With the use of AI, this noice can be reduced and with this, low-dose CT imaging can be used for the predicition of calcification.

Question 13

For what can 3D convolutioan neural networks be used regarding a CT image of the heart?

Answer 13

Automatic segmentation of different structures of the heart to measure e.g. volume measurements and ejection fraction.

Question 14

Name examples of AI applications.

Answer 14

• AI-model for stent-graft segmentation
• Calcium scoring in cardiac CT
• Lumbar artery detection (source of type 2 endoleaks and thus classification of endoleaks)
• Whole-heart segmentation
• Disease classification
• Mortality prediction

Question 15