AI in the hospital

Question 1

difference between artificial intelligence, machine learning, and deep learning

Answer 1

1) AI: mimicking the intelligence or behavioural pattern of humans or any other living entity

2) ML: computer learns from data, without using a complex set of different rules. Approach based on training a model from datasets.

3) DL: perform machine learning inspired by our brain’s own network of neurons

Question 2

what can AI do for healthcare?

Answer 2

• improve quality of care and population health outcomes while reducing costs
• improve diagnostic accuracy
• enhance treatment effectivenesss
• optimize healthcare delivery systems
• model and prevent spread of hospital acquired infections

Question 3

9 areas of application of AI in healthcare

Answer 3

1) diagnostics and imaging (radiology/pathology)
2) predictive analysis (disease, patient outcome)
3) personalized medicine / genomic analysis
4) drug discovery and development (clinical trials)
5) virtual health assistants (chatbots, remote monitoring)
6) administrative efficiency (electronic health records)
7) robotic surgery (precision, minimally invasive)
8) mental health
9) public health (epidemic prediction)

Question 4

what are 3 things we need to ensure when designing an AI?

Answer 4

effectiveness, usability, ethical considerations

Question 5

10 key elements when designing an AI

Answer 5

1) problem definition
2) data collection and preparation (cleaning, correct format)
3) feature engineering -> find features that will be inputs to the AI model
4) algorithm selection
5) model training -> maximize performance and minimize errors
6) evaluation metrics
7) validation and testing (using separate data)
8) deployment and integration
9) monitoring and maintenance
10) ethical and legal considerations (bias, fairness, privacy, transparency)

Question 6

what info can be taken from an MRI to detect multiple sclerosis? What is the gold standard now and what issues could AI help solve?

Answer 6

white/gray matter lesion identification and quantification

Gold standard=manual lesion segmentation -> time consuming and expert variability -> use AI

Question 7

two types of AI to detect the brain matter lesions

Answer 7

1) supervised = rely on manually labeled training set and aim at learning a function that maps the input to the desired output

2) unsupervised = do not require manual annotations as they are based on generative models that rely on modeling the MRI intensity values of different brain tissues and lesions

Question 8

how is the number of lesions correlated to the EDSS?

Answer 8

more lesions when higher EDSS (expanded disability status scale)

Question 9

what is the U-Net neural network architecture? Does it need a small or big training dataset?
Is it supervised or unsupervised?

Answer 9

type of convolutional neural network developed for biomedical image segmentation, needs a small training data set

It is supervised (deep learning)

Contracting path (-> better resolution) and then an expansive path

Question 10

steps of the neural network implementation

Answer 10

1) preprocessing (coregistration between the two types of MRI images)

2) training: all the lesions are samples, regarding of size

3) data augmentation (to avoid overfitting)

4) evaluation strategies and metrics

Question 11

AI to generate synthetic data: name of the network, general concept

Answer 11

Generative adversial neural network (GAN): composed of a generator model and a discriminator model (classifies generated examples as real or fake)

Question 12

what are the three loss functions/types? (GAN model)

Answer 12

• L1 loss (before discriminator)
• adversarial loss (generator fooling the discriminator)
• perceptual loss

Question 13

AI to detect changes on white/gray matter lesions: sueprvised or unsupervised? algorithm flowchart?

Answer 13

unsupervised

1) preprocessing
2) image subtraction and lesion segmentation algorithm
3) threshold

Question 14

what are the 4 types of lesions that can be detected with the unsupervised model?

Answer 14

1) new in TP2 compared to TP1
2) enlarged
3) shrunken
4) stable (lesion not in the above criteria)

Question 15

What are three common myths in building and translating AI models into health care?

Answer 15

1) more data is all you need for a better model -> no, it’s also about data quality

2) an accurate model is all you need for a useful product -> human centered approach is also needed

3) good product is sufficient for clinical impact -> implementation and healthcare economic researcj is critical