Explore a comprehensive framework for developing interpretable diagnostic and prognostic machine learning models using digitized histopathology slides in this 52-minute conference talk by Max Lu from Stanford University. Learn about a scalable method that doesn't require manual annotation of regions of interest and can be applied to tens of thousands of samples. Discover applications ranging from cancer subtyping and prognosis to predicting primary origins of metastatic tumors. Gain insights into weakly-supervised, large-scale computational pathology techniques, including multiple instance learning, attention pooling, and data efficiency. Examine benchmarks, attention scores, and interactive demos showcasing the framework's capabilities in cell phone microscopy, prognosis, and primary origin prediction of metastatic tumors.
Weakly-Supervised, Large-Scale Computational Pathology for Diagnosis and Prognosis - Max Lu
Stanford University via YouTube
Overview
Syllabus
Introduction
Welcome
Background
Example
General workflow
Can we train accurate diagnostic or problem prognostic models
The same label assumption
Multiple instance learning
Data efficiency
Recap
Framework
Segmentation
Embedding
Attention pooling
Summary
Benchmarks
Attention scores
Cell phone microscopy
Results
Summarize
Code
Prognosis
Primary origins of ceps
Study design
Classification
Heatmaps
Interactive demo
Attention heating map
Dummy tool
High certainty diagnosis
Differential diagnosis
Thank you
Which regions in the slide will contribute
Can the primary originate from one single primary
Is the morphology more nuanced
Clustering
Outro
Taught by
Stanford MedAI
