Explore few-shot learning techniques for chest X-ray diagnosis in this Stanford University seminar presented by Dr. Angshuman Paul. Delve into two novel methods: a discriminative ensemble trained on clinical images and a model utilizing both scientific literature and unlabeled clinical chest X-rays. Compare these approaches to existing few-shot learning methods and understand their superior performance. Gain insights into the challenges of few-shot learning in radiology image analysis and learn about ensemble models, bootstrap sampling, subspace sampling, and meta-learning frameworks. Examine the experimental results, including F1 scores and utility of design, and consider future directions in this field. Engage with the speaker's expertise in machine learning, medical imaging, and computer vision during the interactive Q&A session following the presentation.
Overview
Syllabus
Introduction
Outline
Fewshot Learning
Fewshot Challenges
Ensemble Models
Two Methods for Chest Xray Diagnosis
Ensemble Learning
Bootstrap sampling
Projecting bootstrap samples
Subspace sampling
Winner subspace
Subspace dimension
Clusterbased representation
Hidden space representation
Weighted voting
Query input
Process Pipeline
Auto Encoder Ensemble
Class Levels
Metalearning Framework
Experiments
Combinations
Training Data
Results
F1 Scores
Utility of Design
Conclusion
Questions
Future Sector
Classification Pipeline
Initial Training
Loss Function
Pseudo Levels
Retraining
Performance
Result
Taught by
Stanford MedAI
