Artificial Intelligence in Digital Pathology Flashcards
What does artificial intelligence aim to teach machines?
Decision making Reasoning Computer Vision Natural language processing Knowledge representation Motion and manipulation
What are two types of AI?
Artificial general intelligence:
Perform across several domains of human intelligence
Turing test
Narrow (weak) AI:
Perform single task as well as humans
Alpha go
What are subsets of AI?
Machine learning:
Algorithms that learn to act from data without being explicitly programmed
Deep learning:
Subset of machine learning based on neural network models
What does machine learning use?
Data to detect patterns
“a set of methods that can automatically detect patterns in data, and then use the . . . patterns to predict future data, or to perform other kinds of decision making. . . ” – Kevin Murphy
How does traditional programme differ to machine learning?
In traditional programming input and program lead to computation and result
In machine learning input and a desired result lead to computation and result
What three categorised is machine learning divided into?
Supervised: Data modelled using label . Used to cluster data
Unsupervised: Data modelled without labels.
Used to cluster data
Reinforcement: Learn from experience. Model uses trial and error to learn from experience. A reward function is used to measure how good each action taken was
Supervised learning techniques include classification and regression
How do both differ?
Classification-
Discrete or categorical labels
Separate observations into categories
Classifier finds mapping between features and classes (categories) that can can predict category for future observations
Logistic regression and random forests common classifiers
Regression- Numerical or continuous output variables Explanatory variables (features) Try to predict a dependent variable Finds function that maps between dependent variable and explanatory variables (features) which can be used to predict a real value for future observations
What does deep learning use?
Neural networks based on biological neurons networks
- Deep learning is based upon what?
2. Why is this powerful?
- Computational networks which are loosely based upon biological neural networks.
- As known as universal function approximates meaning can model complex functions.
What are features of the deep learning model?
Input layer: features for each observation enter the network
Hidden layers: the neurons perform feature extraction and learning task
Output = prediction
In deep learning what is making decisions about its inputs based on the weights it tries to learn during training?
A single neuron
In supervised deep learning a loss function evaluates…
what predictions adjusts the parameters of the network based
What are the two phases within neural networks?
Training and Inference.
During training learning most important features and weights to produce accurate prediction of the weights. Weights extremely important as informing of which features are significant in forming final prediction. This is achieved through an iterative process using two segments: Forward pass and backward pass. Forward pass: Learn which features are important through process using calculation for each neuron to form a decision about its input based upon the weights it tries to learn during training, use features to form a prediction at end of network. Backward pass: Compare prediction to ground truth label using function we defined and adjust weights in the network using an optimisation algorithm - this adjusts the weights proportional to the loss calculated such that can improve on this in next iteration.
Inference: Only single forward pass here whereby weights are static and set at level learnt during training phase. Use this to predict data, extract learnt features and form prediction at end of network
Computer vision is the area teaching machines to understand visual data
What are 4 main tasks often performed using computer vision?
Classification– Correctly label image
Classification and Localization- Correctly label and locate image
Object Detection- Localising all objects of interest in the image
Instance Segmentation- Classification task attempting to apply to all images
In computer vision image date is?
How does this differ to machine learning?
Unstructured and features are learned or engineered from raw pixel values
In machine learning must engineer features manually, whereas here this is not needed
Computer vision in histopathology focuses upon what tasks?
Classification and Segmentation
Computer vision is a form of deep learning
True or false
True
What are applications of AI in digital pathology?
Detect or grade cancer
Identify metastases
Prognosis prediction
Personalised care plans
Aid pathologist in decision making which could lead to
Improved efficiency
Improved workflow
More advanced diagnostics
Whole-slide scanners are used digitize glass slides containing tissue specimens at high resolution.
Why is this beneficial?
High resolution- Typical resolutions of 100,000×100,000
Images are stored at different magnifications – generally up to 40x
How do pathologists annotate whole slider scanner images (WSIs)?
What is a limitation of this?
How can this be overcome and what are the advantages of this?
Manually annotate images using programs designed to deal with these images – Qupath
Annotating single lymph node extremely time consuming – estimated max 30-40 lymph nodes a day!
Given WSIs are a large source of information images can be analysed using AI techniques
Automation beneficial because
- Free up pathologist time
- Improve accuracy, reproducibility and reduce potential errors
- Improve inter-pathologist agreement
Automation order of magnitude quicker once models are trained!
How is machine learning used to analyse whole slider scanner images (WSIs)?
Split the WSI images into a train and test dataset
Automate annotation of lymph nodes for histological features using a deep learning segmentation pipeline
Compare network predictions with pathologist ground truth annotation
Stitch patches back together to create entire segmented whole slide image
What is the Camelyon16/17 challenge?
Goal: evaluate algorithms for automated detection of metastases in H&E whole-slide images of lymph node sections.
400 whole-slide images (WSIs) of sentinel lymph nodes
1st Task:
Whole slide level prediction
Binary classification problem
Assign metastatic or not metastatic to each image
2nd Task
Find metastasis location
Segmentation problem
Assign metastatic or not metastatic to each pixel
What are the current challenges of AI and digital pathology?
Lack of labelled data
AI requires large set of training images – ideally labelled
Extreme variability
Large variation in tissue type
Extreme polymorphism
Dimensionality obstacle
Gigapixel digital images
Memory intensive
Down-sampling my result in loss of crucial information
Computationally intensive
Large data requirements
GPUs required for training (otherwise patience required!)
Narrow AI
Deep NNs are belong to weak AI
Train multiple AI solutions for different tasks
Lack of transparency Black box No obvious rationale behind decisions Unacceptable in medical discipline Decision making process behind diagnosis needs to be clear Transparency key for regulatory approval
Adversarial attacks
DNNs can be easily fooled
Noise or artifacts skew output – mistakenly diagnose as cancer or vice versa
AI in practice Buy in from pathologists Ease of use Scalable Trust understandable
What are ethical considerations in AI ?
Bias
Natural bias in society can bias data
Biased data can lead to biased decisions and results
Shared prosperity in AI
Automation revolution
Job displacement
Effects on skill requirements
Transparency
Builds public trust in system
Important to explain why decisions are being made
Safety certification
Data protection
protection of sensitive or individual data
Security
Can the AI be hacked or used for malicious
purposes
Accountability
Erroneous decisions
Who is responsible if machines make mistake or mis-diagnosis
Cathy O’Neil coined weapons of math destruction
What does this refer to?
ML models have the capacity to create great damage to
society.
Weapons of math destruction models:
Opacity, Black box models: cannot easily be interrogated. Especially by the subjects.
Scale: Can the effects of the model extend beyond the decision that it helps to make.
Damage: The model causes damage. Could cause harm
to people, especially members of vulnerable groups.
