AI in Medical Diagnosis - Module 1 Flashcards
Advances in the last decade have largely been due to three factors
- Maturation of deep learning
- Compute power via GPUs
- Open-sourcing of large labeled datasets
Clinical tasks suitable for Computer Vision
Screening Diagnosis Detecting conditions Outcome prediction Segmentation of pathologies Monitoring disease Clinical research
Object classification
Refers to identifying the type of object in an image
Object localization
Refers to the localization of an object in a image
Object detection in computer vision
Refers to identifying the type of object and location of the object
What is the key technique that leveraged CV
Convolutional neural networks
Characteristic of the mechanism of a CNN
Hardcodes translational invariance
Traslational invariance is
The system produce the same response regardless of how the input is shifted
Image registration
identifying corresponding points across similar
images
image retrieval
finding similar images
Data points in CV
Images
Data labels in CV
Object classes
Why transfer learning has been critical in medical AI
Given the sparcity and access difficulties of medical data
3 Strategies to reduce costs associated with collecting and labeling data
- Data augmentation
- Generative adversarial networks
- Crowd sourcing
Self supervised learning means
Implicit labels are extracted from data points and used to train algorithms
Federated learning means
Centralized algorithms can be trained on distributed data
Multimodal learning
Combining vision with other modalities
Useful in detecting adverse clinical events
Activity recognition
Live scene undestanding
Challenges in DL based computer vision with medical imagery
Massive images
Variance on technique of digitalizing images
3D images
Multiple instance learning (MIL)
Is a type of supervised learning. Instead of receiving a set of instances which are individually labeled, the learner receives a set of labeled bags, each containing many instances.
How to deal with non standarized data collection
Integrating computer vision in existing physical systems
CV in gastroenterology
Esophageal cancer detection Lesion detection (gastric cancer) Lesion diagnosis (h pilory)
CV in radiology
Analysis of brain (stroke)
Nodule detection
Region segmentation (ventricular or coronary)
New uses of CV in radiology
Image reconstruction and enhancement
Automated report generation
Temporal tracking
CV in cardiology
Coronary artery segmentation
Echocardiography (hypertrophic cardiomyopathy, cardiac amyloid and pulmonary hypertension)
CV in Pathology potentials
- Overcome limitations of human vision and cognition
- Develop new signatures of disease and therapy
- Combination of pathology with radiological, genomic and proteomic measures to improve prognosis
CV in dermatology
Lesion specific differential diagnostics
Finding concerning lesions against bening
Track lesion over time
CV in Opthalmology
Diagnosis of diabetic retinopathy
Macular edema
Glaucoma
Non ocular complications
CV in surgical video
- Increase performance through real time context awareness
- Skills assesments
- Training
GOALS
Global
Operative Assessment of Laparoscopic Skills criteria
Ambient intelligence
Computer vision coupled with sensors
and video streams enables a number of safety applications in clinical
and home settings, enabling healthcare providers to scale their
ability to monitor patients.
Key considerations when applying ML in healthcare
- Assessment of data
- Planning model limitations
- Community participation
- Trust building
Methods to remove individual level bias
Expert discussion
Label adjudication
Method to remove population level bias
Missing data supplements
Distributional shifts
Multi institutional evaluation
Multi task learning (eg. multicancer detection rather than one type of cancer)
Assess demographic performance w saliency maps
Model limitations planning
Developing confidence intervals
Explainability
To increase community participation
Side by side deployment with existing workflows
AI interpretability
why specific factors about
the patient or environment lead them to their predictions
Federated learning
Federated Learning enables mobile phones to collaboratively learn a shared prediction model while keeping all the training data on device
inherent explainability
For machine learning models for which the input data are of limited complexity and clearly understandable (eg. linear regression)
post-hoc explainability
the data and models are too
complex and high-dimensional to be easily understood; they cannot be explained by a simple relationship dissect the model’s decision making procedure
Heat map or Saliency maps
highlight how much each
region of the image contributed to a given decision and are illustrative because they provide a simple means of understanding some of the limitations of post-hoc explainability techniques
adversarial attack
when precise modifications are made to the input that
substantially alter the model’s predictions
