As the interests of radiology residents and consultants in AI keeps increasing, there has been a lot of demand to explain the nuances of AI to enhance knowledge. Radiologists looking at and decoding AI need to figure out the key roles they are likely to play in this journey.
This article will explain the various possible roles you can play in interacting with AI, and give a list of suggested reading for you to take the initial steps to understand more on AI. Radiologists will be interacting with AI in one of the following ways, led by the early adopters.
I] The “Early Adopters” will interact with the AI offerings considering the have the following roles and responsibilities.
- As key strategic decision makers on choosing the right AI products and solutions for the radiology departments, their clinics and hospitals.
- As hands on Data Annotation Experts laying the blueprint on Data Annotation and Data Validation Activities.
- As research leaders through their domain specific knowledge required for development of AI products and solutions as a part of their research labs at hospitals.
II] The “Late Adopters” who will interact and get involved with AI through learnings of peers.
The three key reasons why AI discussions will happen in the radiology department will be to ensure how it can increase efficiency, minimize errors and reduce burnouts.
A] Early adopters like those in key strategic decision-making roles are needed to ensure that AI solutions which they bring in can address these expectations.
Early Adopters are likely to interact with a number of point AI solutions. Here the radiologists need to decode how AI will actually contribute and impact their day to day activities. There are three ways that AI is likely to come into the Radiology Department.
- On the devices provided by vendors where typically as an imaging expert you don’t have much say as you simply adapt and use it.
- These solutions could be cloud based and which may need you to push your data to a cloud-based server for analysis.
- These may reside on the hospital PACS.
Imaging experts must understand that each of these options can have their own peculiar advantages and disadvantages.
Imaging experts will have to decipher the differences between various point solutions available and their key differentiations factors. Radiologists will also need to familiarize themselves how the algorithms work and familiarize themselves with sensitivity, specificity, F1 scores, Area Under Curve and Confidence Intervals at which the scores are derived for algorithms as these parameters will eventually impacts the True and False Positive cases they will encounter for that particular disease condition.
Though these key matrices for algorithm checks and balances are important, they are still not the entire truth. In depth knowledge in terms of actual performance of the solutions by running demo solutions and understanding its performance with hospital dataset will give a clear analysis. Imaging experts playing a role in decision making need to understand that the algorithm is not a one size fit all solution but needs to be fine tunes after understanding the fine balances between the false and true positives. On the brighter side once you find the sweet spot, it has the potential to significantly amplify the workflows.
B] If radiologists want to support the Data Annotation Activities (DAA)
This is a key activity needed for generating good quality algorithms. AI is data hungry. Clean and structured data with correct annotations play an important role and need close coordination between Radiologists and Data Science Team. This boils down to defining the aims and objectives of the DAA and to understanding the outcomes of the end result expected out of the algorithm. DAA is extremely broad terminology and gets defined by deliberating if we are creating a multiclass or binary classification models. DAA for X-ray, Ultrasound, CT and MRI require high skills and quality review so the final solution does not falter, or get compromised or has any bias. DAA by imaging experts need review and quality checks by another team of experts to exclude bias. This also needs to be coorelated with other existing ground truth which has to be custom designed based on the use case getting evolved. Imaging experts also need to understand workings of various on premise and cloud solutions available for DA development, storage and management.
C] If hospitals are trying to develop AI products and solutions as a part of their hospitals and research labs.
This is an ambitious activity and usually carried out with a large team of Data Science and Technology Experts. The goal of these exercises is to create AI algorithms for hospital research activities and may potentially undergo commercialization at a later date. This requires a larger framework of controlled development based on the overall outcomes expected after going through the exercise of analysis of benefit to hospitals or society at large. These activities are usually taken by hospitals or governmental organizations trying to solve problems related to triage for specific disease conditions.
Are radiologists expected to write code and learn languages like Python?
Well this is not a requirement or a must unless you are naturally inclined and want to experiment and learn coding.
AI is now being all pervading and omnipresent, it already exists in the form of Alexa or Siri on your mobiles. For imaging experts it’s already coming in some form of auto segmentation directly integrated to modality. The straightening of the coronary arteries on your cardiac CT, segmentation of the liver in Liver Elastography on MRI scanners and segmentation of cartilage on cartilage post processing tools. It’s already there either as a component of machine learning or deep learning tools. You may also encounter and experience AI in your hospital PACS or vendors may even come with their own Platforms.
Hope you spend time to experiment and learn by interacting with AI solutions and play a major part in integrating it in your facility or practice. The more you interact more it widens your horizons and therefor the opportunity for you to understand AI better.
For additional reading and course material you can simply google your way around and there is enough material to keep you busy for a pretty long time. However, I can suggest a few links and research materials just as a starter kit –
Understanding AI: https://www.tutorialandexample.com/artificial-intelligence-tutorial/
Understanding Quality and Metrics in AI: https://www.sciencedirect.com/science/article/pii/S155372502030204X?via%3Dihub
How does data annotation work – imaging experts can review this blog: https://blogs.nvidia.com/blog/2019/10/17/ai-healthcare-training-data/
AI learning from journals: https://pubs.rsna.org/journal/ai
For the enthusiast: https://www.coursera.org/learn/ai-for-everyone
For the relaxed reader who likes reading stories you can go through the book- The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World.
I am also sharing some links of some interesting pubmed articles for residents and those interested in understanding more details on AI:
Artificial intelligence in radiology: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6268174/
Deep learning in radiology: An overview of the concepts and a survey of the state of the art with focus on MR: https://pubmed.ncbi.nlm.nih.gov/30575178/
Artificial intelligence in image analysis-fundamentals and new developments: https://pubmed.ncbi.nlm.nih.gov/32789670/
Ethical considerations for artificial intelligence: an overview of the current radiology landscape: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7490024/
How Cognitive Machines Can Augment Medical Imaging: https://www.ajronline.org/doi/10.2214/AJR.18.19914
Hope this helps!
– Dr Amit Kharat, MBBS, DNB, PhD
Fellow of Indian College of Radiology and Imaging
Professor of Radiology and Cofounder, http://www.Deeptek.ai