Building AI models for healthcare (ML Tech Talks)
Вставка
- Опубліковано 13 чер 2024
- In this session of Machine Learning Tech Talks, Product Manager Lily Peng will discuss the three common myths in building AI models for healthcare.
Chapters:
0:00 - Introduction
1:48 - Myth #1: More data is all you need for a better model
6:58 - Myth #2: An accurate model is all you need for a useful product
9:15 - Myth #3: A good product is sufficient for clinical impact
12:19 - Conversation with Kira Whitehouse, Software Engineer
34:48 - Conversation with Scott McKinney, Software Engineer
Resources:
Deep Learning for Detection of Diabetic Eye Disease: Gulshan et al, Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA 2016 → goo.gle/3gVhTxs
A major milestone for the treatment of eye disease De Fauw et al, Clinically applicable deep learning for diagnosis and referral in retinal disease. Nature Medicine September 2018 → goo.gle/35Sfs9C
Assessing Cardiovascular Risk Factors with Computer Vision. Poplin et al, Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature Biomedical Engineering. March 2018 → goo.gle/3qkg01I
Improving the Effectiveness of Diabetic Retinopathy Models: Krause et al, Grader Variability and the Importance of Reference Standards for Evaluating Machine Learning Models for Diabetic Retinopathy. Ophthalmology August 2018 → goo.gle/3gR8d8n
Deep learning versus human graders for classifying diabetic retinopathy severity in a nationwide screening program. Raumviboonsuk et al. NPJ Digital Medicine. April 2019 → goo.gle/2SmyXUO
Healthcare AI systems that put people at the center: Beede et al, A Human-Centered Evaluation of a Deep Learning System Deployed in Clinics for the Detection of Diabetic Retinopathy. CHI '20 April 2020 → goo.gle/3ja6TyP
Artificial intelligence for teleophthalmology-based diabetic retinopathy screening in a national programme: an economic analysis modelling study. MScPH, Yuchen Xie, Quang D. Nguyen BEng, Haslina Hamzah BSc, Gilbert Lim, Valentina Bellemo MSc, Dinesh V. Gunasekeran MBBS, Michelle Y. Yip, et al. The Lancet → goo.gle/3zVec3q
Catch more ML Tech Talks → goo.gle/ml-tech-talks
Subscribe to TensorFlow → goo.gle/TensorFlow - Наука та технологія
It would be great to have more videos about the applications of AI in Healthcare
This is so inspiring . I am a medical student too and studying ai on my own . Thankful for coming across this video
same here
What an amazing conversation! Thank you! Interesting point about the dust on the camera, sometimes a simple rubber air dust blower will do 90% of the time before any shots taken (a trick I learnt from photographers). Many clean their lenses with Isopropyl alcohol, and good solvent as it is, will wipe that coating off and damage the image quality later.
What an amazing conversation. Thank you for sharing!
That is great !!! Not only in the AI field but Translational field as well. T0:Basic Science -> T1:Translation to Humans -> T2:Translation to Patients -> T3:Translation to Clinical Practice -> T4:Translation to Populations.
Being a Medical AI student, this lecture has helped me a lot to think about things from different perspectives, especially related to building models/solutions for healthcare.
Thank you
Great video & conversation. Watched from beginning to end, couldn't stop. Looking forward to more content like this.
this was nice and refreshing in taking a realistic approach and being fairly open about the challenges and downplaying the hype - thanks
Simply wonderful. Thank you!
I really liked the 3 rules from Scott. They helped me a lot in my research in medical image analysis.
Love this video is very informative. Clearly one needs familiarity with healthcare in order to build good use cases since business objectives could be different than other businesses .
Great video! I do like to hear more and more about data quality over data quantity, sounds promising. Btw, did you guys where trying to make a reference to Transformers while saying "is all you need" multiple times? :D
A great talk. I appreciate all the sharing in this video
Great discussion! Great insights! I was trying to train a neural network to compete in the Kaggle competition for identifying pneumothoracies. I got sidetracked into trying to classify different pneumothoracies (those with mediastianal shift ... tension pneumothorax, those with atelectasis (large pneumothoracies), those with a chest tube, and those hard to see pneumothoracies). Never finished a model that would be competitive. Encountered the same thing with Kaggle competition on intracerebral hemorrhage. In those cases the issues were mainly with patients who already had surgical changes in the skull. There are simple class data points that could be incorporated into a model ... patient has a chest tube, patient is post op craniotomy, patient is a battlefield casualty.
A valuable asset in engineering a process are the technicians. They may not have the training of the medical expert, but they do have training in spotting artifacts and other problems with image quality.
I would think if you build a system to detect diabetic retinopathy, you would want to include the detection of cataracts (or corneal diseases). Producing a quantitative estimate of blurring due to cataracts would give a practical measure of the accuracy of diagnosis of retinopathy, and identify those who might benefit from other medical interventions.
Simple but very informative to drive AI toward useful productivity.
Great explanation really easy to understand!
Thank you for this amazingly insightful vlog that illustrates the real life challenges of ML and the secret sauce is persistence, skepticism and the ability to work together for a solution. That is why, attitude and culture is critical if we want to create useful models that overcome over innate biases and misconceptions. Grateful for your insights.
Perharps, you could add image enhancement to the images to filter some of the lesions that appear as a result of dust on the lens etc
This was superb, thanks for all the insights and pro tips :)
Great lessons that we can't find in school.
..thank you for a such an informative video very clearly put together, I have been attending IBM EDT courses..specifically the AI practitoner course..the framework addresses similar issues holistically.., 25 years in electronics manufacturing ..I thought I seen it all ..EDT opened my eyes wider and let in more knowledge....I am passionate and know I can contribute in AI healthcare ..but dont know how to get in
تحياتي الخالصة من الجزائر thank you very mutch
IN Health Care in order to build a "good" dataset it's needed a sound background in epidemiology and in medical statistics.
More videos about the applications of AI in Healthcare please
Very relevant discussion! Great video!
Thanks. Its very resourceful. I am curios if pre-trained model is available.
I'm gonna hit the like button 10 times. Machine learning is the best.
Can you tell why the performance of the model plateaued at a certain number of observations [4:20]?
It's an interesting topic if you are in healthcare. But, out of curiosity how long does it normally take for an optometrist to view the image and make the diagnoses? I couldn't imagine it being that long as in maybe 10-15 seconds? I suppose this is more useful if you only have a tech on hand and not a optometrist.
thank u so mush, where can i found the data to train my algo ?
Thank you!
46:20 general rule to build better models: be skeptical. Question easy good results (a bug is there probably) since good model are very unlikely to be built easily.
38:00 Lesson 2: question construction of data sets. Interact with data curator to avoid model cheating. Example of tuberculosis identification from chest images.
Tensorflow ❤
It seems that most of the problems mentioned regarding the design of the Product are related to not having a standarized Design for Machine Learning best practice.
class!!
Great video, thanks!
Are there any services where you can find highly qualified, for example, doctors, for data labelling?
35:10 great 3 lessons for building healthcare models, from an expert:
1. Look at the data. Browse them even not an expert. At 36:30 Funny story of the model detecting circled notes on images due to data contamination.
the dream was having neural networks decide the trash filter properties, if something is contaminated (cheat) against error. but its also a error if it worked wrong.
THIS IS cOOL IT wILL REALLY HELP
and where is the part when you start to build ? I only see discussion
lost my time
23:04 “Why do they don’t come back after they got diagnosed for follow up?” Simple - You live in a country without general healthcare where treating an illness, getting there, staying healthy is a financial burden. Talking about Thailand and the United State of America. 🤷🏻♂️