Pranav Rajpurkar
Stanford University

April 17, 2019
The use of algorithms in clinical care demands a very high performance level for accurate detection and classification of disease. Deep learning (DL) offers a powerful toolkit necessary to handle the complex variations present in medical data, which traditional statistical or machine learning approaches have historically been unable to capture. In this talk, I will describe the challenges and approaches for the development of high-performance DL algorithms and curation of datasets for problems in diagnostic radiology and cardiology. I will also discuss the use of these algorithms as diagnostic support tools for clinicians, and challenges for the potential translation of these algorithms from the lab setting to clinical practice.

View the full playlist: [ Ссылка ]

0:00 Introduction
0:37 Traditional Model of Clinical
2:42 Diagnostic Procedure with Al
4:02 Information Gathering Step
5:44 Diagnostic Testing with Al
6:49 Machine Learning Framework
8:54 Next Paradigm shift?
10:08 Arrhythmia detection
10:49 Future of continuous monitoring
11:56 Holter Monitor
12:27 Amount of data capture
15:16 Automated Detection Challenges
18:20 Previous Approaches
19:10 Setup
20:29 Network Architecture
20:50 Residual Networks
21:42 Wide ResNets
23:05 Dataset - Test Set
30:48 Medical Errors
36:36 Chest Radiograph Interpretation
37:02 Chest X-ray exam
37:19 Detecting Abnormalities
37:32 X-ray findings of pneumonia
37:46 Detecting Pneumonia
38:30 DenseNets
39:17 Dataset - Train Set
39:52 Evaluation -- Limitations
41:49 Class Activation Maps
42:07 Interpretations
42:39 Future of diagnostic access
46:53 Diagnosis Clinical Decision Support
48:34 Knee MR
49:26 Input
50:38 Dataset Training on
51:07 Interpretability
53:28 External Validation
55:44 Diagnostic Future with Al

свернуть