Andrew Hartz MD
Michael Romano M4 MD/PhD program
Jonathan Scalera MD
Nothing Artificial About This Man
Look deep into nature and then you will understand everything better Einstein
Overview
Radiology residents are taught the theories and physics concepts supporting MRI, CT, radiography, and other medical imaging technologies. AI is rapidly becoming the newest imaging technology used in radiology practices. The goal of this project is to design a curriculum for educating radiology residents about AI.
The current state of AI in Radiology: A Survey
Image Enhancement
- Decreasing contrast dose: see
Use of artificial intelligence in computed tomography dose optimisation
- Decreasing exposure time
- Increasing image quality (see below)
- Lower magnet strengths: see Enhancing magnetic resonance imaging-driven Alzheimer’s disease classification performance using generative adversarial learning
- Mobile MRI scanners: see Hyperfine Swoop
Triaging
- Decrease time-to-read
AI technology can be used to optimize workflow by assisting in prioritizing the exams with the most concerning findings. Typical radiology practices use date and time of exam for prioritizing studies. AI can help detect concerning pathology automatically before the radiologist opens the study. Please see, for example: Automated deep-neural-network surveillance of cranial images for acute neurologic events
Segmentation
- Decrease time spent by radiologist on manually segmenting. For examples of papers with corresponding code, please see: 3D Medical Imaging Segmentation
Third Reader
- Increasing speed
Similar to humans, the performance of deep learning networks continue to improve as it trains on more data.
Natural Language Processing
- More accuracy with transcribing
- More accuracy with translating exam indication
- Improved management based on comprehension of written reports
- Please see a nice review here: Natural Language Processing in Radiology: A Systematic Review
Research
- Finding new areas of research
- Curating articles, for example: Scaling up data curation using deep learning: An application to literature triage in genomic variation resources
- Improving healthcare systems and limitations
Deep learning algorithms are able to identify patterns outside of human perception. New areas of interest can be discovered using deep learning.
AI has the potential to improve the workflow in radiology. Increased complexity of AI application can also hinder the workflow. Furthermore, there is the increased costs associated with implementing AI systems that must be considered.
Curriculum
Theory
-
Note: the following links are to chapters from Goodfellow et al. (2016).
- The Retina and The Brain
- Probability. Another excellent introduction to statistical methods is All of Statistics
- Dimensionality
- Vector Manipulation
- Gradient Descent (see sections 4.1-4.3)
- Back-propagation (see section 6.5)
- Convolutions (see section 9.1)
- Neural Networks (see section 6.1)
Deep learning was inspired by the human neural network of the brain, and its ability to translate visual information obtained by the retina.
Problem-based learning
- How to gather important findings from a chest radiograph?
- Chest X-Ray Reader/Chester AI
- How to create a trained neural network using a database of thousands of radiographs?
- Building a trained neural network
Limitations
- Adversarial Examples
- “Understandable AI”. For a complete understanding of outstanding issues here, please see: Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI
Education
Books
- MIT – Deep Learning Book
- Introduction to Statistical Learning (in R)
- Elements of Statistical Learning. A more advanced and mathematically-heavy text than Introduction to Statistical Learning.
Projects
BMC Radiology – Chest X-ray Reader
https://ai.thecommonvein.net/chest-x-ray-reader/
Created by Andrew Hartz, MD and Michael Romano, MS4 MD/PhD Program
Adapted from Chester the AI Radiology Assistant (https://mlmed.org/tools/xray/)
Use with Google Chrome (JavaScript enabled) for best results. Allow the program to load (up to 2 minutes) after uploading any chest x-ray image. Results will be shown on a probability table and saliency map.
Research
AI in Medicine
- JAMA Machine Learning
- Doctor Penguin (AI in Healthcare)
- AI in Healthcare
AI in Radiology
- RSNA Radiology: AI (Journal)
- RSNA AI Resources
- AI Imaging Business News
- Dartmouth’s AI-RADS Lectures (https://www.sciencedirect.com/science/article/pii/S1076633220305560):
- RFS Artificial Intelligence Journal Club
- Stanford – Center for Artificial Intelligence in Medicine & Imaging
- Mount Sinai Hospital AISinai
- NYU – CAI2R
- MGH
- BWH
- UVA AI lab
- UW
- Duke
Boston University Resources