MedAI - Training Medical Image Segmentation Models with Less Labeled Data - Sarah Hooper

Explore a comprehensive lecture on training medical image segmentation models with reduced labeled data requirements. Delve into Sarah Hooper's research at Stanford University, focusing on a semi-supervised method that significantly decreases the need for extensive labeled datasets in neural network training for medical image segmentation. Learn about the application of this technique to cardiac magnetic resonance (CMR) segmentation, its impact on deriving cardiac functional biomarkers, and the potential for making quality healthcare more accessible. Gain insights into the two-step process involving data augmentation, traditional supervision, and self-supervised learning, as well as the evaluation of labeling reduction, error modes, and generalization capabilities of the proposed model.

Intro
Many use cases for deep-learning based medical image segmentation
Goal: develop and validate methods to use mostly unlabeled data to train segmentation networks.
Overview Inputs: labeled data. S, and labeled data, Our approach two-step process using data augmentation with traditional supervision, self supervised learning and
Supervised loss: learn from the labeled data
Self-supervised loss: learn from the unlabeled data
Step 1: train initial segmentation network
Main evaluation questions
Tasks and evaluation metrics
Labeling reduction
Step 2: pseudo-label and retrain
Visualizations
Error modes
Biomarker evaluation
Generalization
Strengths