Radiologist-in-the-loop AI
Emphasizing the synergy of radiologists' and machines' intelligence to enhance the performance of radiological AI systems!*
* This will be a tutorial taught at MICCAI2023
Explore the power of collaboration between radiologists and machine learning researchers
Learning Objectives
- Design a radiologist-in-the-loop AI system architecture
- Use various active learning strategies to sample data for radiologists’ annotation
- Create labeling instructions for knowledge transfer
- Apply the new approach to a real-world problem of liver view classification in ultrasound images (i.e., images with liver vs. without liver)
What You Will Learn
- What is radiologist-in-the-loop AI?
Describes how radiologists and machine learning algorithms interact to solve radiological AI problems.
- Introduction to Active Learning (AL)
Introduces the concept of AL and how it differs from a random approach by using a systematic method for selecting data to be annotated.
- Uncertainty sampling
Describes strategies for identifying images that are near a decision boundary in the current ML model.
- Diversity sampling
Describes strategies for identifying unlabeled images that are underrepresented in the current ML model.
- PyTorch implementation of the sampling strategies
Demonstrates how to implement the image sampling strategies and apply them to real-world problems.
- Prepare clear labeling instructions
Explains how knowledge-transfer can be achieved.
- Liver view classification
How the concepts learned above can be used to perform liver view classification in ultrasound images (i.e., images with liver vs. images without liver).
Instructor
Abder-Rahman Ali, PhD
Research Fellow
Harvard Medical School/Massachusetts General Hospital
aali25 [at] mgh.harvard.edu
