New in
@JAMAOphth
! In this brief report, we show that a generalist vision-language model (VLM) may have limited capabilities in performing tasks related to the interpretation of OCT scans.
@pearsekeane
@reena_cho
Match results are in and I’m going to be a retina surgeon! So excited to be joining the
@ClevelandClinic
starting July 2024! It’s an honour to continue my father’s legacy in retina and to join the Cole Eye Institute.
A heartfelt thank you to my wife and family, my mentors at
Beyond thrilled to be joining
@pearsekeane
and his team at
@Moorfields
@UCLeye
for a year-long research fellowship.
Their pioneering work in AI for
#ophthalmology
is truly inspiring, and I am honoured to have the opportunity to learn from them.
This is the first step in my
#Art
,
#ophthalmology
and
#artificialIntelligence
.
Have you heard of DALL·E (
@OpenAI
)? It's a neural network that creates images from text captions.
I created paintings depicting
#cataract
surgery in the style of 4 famous artists: Picasso🇪🇸, Klimt🇦🇹, Munch🇳🇴, and Basquiat🇺🇸.
ChatGPT can write code and answer
#USMLE
questions.
But how does it perform in
#ophthalmology
?
We tested
#ChatGPT
on questions from the OKAP - a high-stakes exam taken annually by residents in the US and Canada.
Here’s what we found:
💙 Feeling grateful as I reflect on the last 5 years with the Dept. of
#ophthalmology
at
@UMontreal
. From the moment I joined, I found a very supportive environment for my ideas and projects.
👨⚕️ As my residency draws to a close, I can't thank my professors enough. Their
Happy to share our recent work on
#GPT4
in
#ophthalmology
👁️📚. Published in
@BMJ_Ophth
's collection on Generative AI 🤖📄. Evaluating LLMs is a minefield 🔍💥, and we're aware there's much more to uncover 🧐 and learn 📖. Key insights below! 👇
🎓 Honoured to have delivered the Valedictorian speech at the MD & Residency Graduations of
@UMontreal
! Touched on
#AI
in medicine and highlighted the importance of holding on to our humanity. I also shared 5 practical tips for recent graduates:
Happy Valentine's Day! 🌹This is a retinal image of Valsalva retinopathy, showing a hemorrhage shaped like a backwards-facing red rose.
The macula helps us appreciate the beauty of the world, like the flawless form of a rose. When the macula is diseased, this experience is
🚀 Apple Vision Pro is here! A VR revolution is coming.
1/ Introducing RetinaVR - our standalone simulator for vitreoretinal training. Immersive, affordable, portable. A step towards democratising surgical education globally.
Details of our pilot work:
Honored to have contributed to this work in
@AAOjournal
on Zellweger Spectrum Disorder. It was a privilege to team up with researchers committed to this rare disease. Hoping this lays a foundation for future therapeutic breakthroughs!
Great work,
@dominicwllmsn
, for leading our project on using AI to enhance clinical trial recruitment in geographic atrophy.
The AI system can shortlist patients by looking through their OCTs for unique imaging inclusion criteria like GA lesion area, foveal involvement, etc.
Preprint alert.
#AI
could aid
#clinicaltrials
recruitment of people with geographic atrophy (GA), which causes irreversible vision loss and affects 5-10 million globally.
The study used anonymised eye data from 300k+ patients at
@Moorfields
🔗
Really honored to receive, for a 2nd consecutive year, the Best Presentation Award at our
@Umontreal
Annual Research Day. As it happens, in 1975 and 1976, my father Samir also won that award 2 years in a row! 🧐
#Ophthalmology
#ArtificialIntelligence
#Retina
📢 Update: We published a preprint in January with some of our results on
#ChatGPT
's accuracy in
#ophthalmology
question-answering. Now, our published paper in
@AAOjournal
offers valuable new lessons. 🤖👁️
🎯 Check out the key takeaways in this thread:
👋🏻 Happy to share my talk on LLMs in ophthalmology from Day 1 of the CCOI 2024 conference. I spoke about their emerging role in medicine and eye care, emphasising their potential and necessary caution. Hope you find this helpful!
🤖 I invite you to join Day 2 of the CCOI 2024
Happy to share our study on
#AI
for the diagnosis of
#trachoma
- the 1st infectious cause of blindness 🌍
Combining
#AutoML
+
#Edge
computing brings the full spectrum of model design, from development to deployment, into the hands of clinicians. This democratizes access to AI.
Excited to spotlight this work led by
@pearsekeane
&
@Yukunzhou19
, published in
@Nature
today. Honoured to have joined this world-class team in July—I'm learning so much every day.🙏🏻
👁 RETFound is the 1st open-source foundation model for retinal imaging—diagnosing eye diseases
#AutoML
democratizes access to
#AI
- allowing clinicians w/o coding expertise to develop
#ML
models
Had the pleasure of sharing my experience at
@EiasSaih
Impossible w/o the trailblazing work done by
@pearsekeane
at
@UCLeye
🙏🏻
Full Video (English cc):
Honored to have been elected Chief Resident of the
@Umontreal
#Ophthalmology
Residency Program! In 1974, my father Samir held the same position then went on to become a VR surgeon. He returned home to Aleppo (Syria) and performed the 1st buckles in the region!
@chumontreal
Check out our proof-of-concept study in
@SciReports
. We built an automated machine learning model for
#cataract
surgery phase classification.
#AutoML
democratizes access to
#AI
- allowing clinicians w/o coding experience to develop high-quality
#ML
models.
Ali Dirani et al
@crchuqc
use deep learning incorporating (HD-OCT) B-scans and clinical features to model visual improvement following macular hole surgery.
Really happy to announce that earlier this June, at the COS annual meeting, I received the Innovation in Retina Research Award for our proposal to improve access to retinal surgery training via low-cost virtual reality simulation.
#Ophthalmology
#Ophthotwitter
#VirtualReality
@kevinelaimd
1st year of residency: focus on BCSCs, esp. cornea, glaucoma, peds, neuro, plastics, retina. Later years, do BCSC Self-Assessment & OphthoQuestions only - easier to do post-OR/clinic. Aim for 20-40 questions daily, doubling on weekends. Make up for missed days. Worked for me!
An honour to meet and hang out with
@mustafasuleyman
- a leader and inspiration in the world of
#AI
. Congrats on the new book 'The Coming Wave'. Looking forward to reading it!
Had so much fun being on this podcast with
@iDrFeng
!
We delved into everything
#AI
: from
#ChatGPT
and
#LLMs
to self-attention, the
#Turing
Test, and multimodality.
🚨 Listen till the end to discover why Anthony Fauci and Joe Rogan also made it into our conversation!
*AI course for clinicians* (in French 🇫🇷)
On Oct. 20 2022, together with the CHUM School of AI in Healthcare, I will be giving an in-person course about automated machine learning.
If you want to use
#AI
but don't know where to start, this course is for you! No coding required.
🎙️ Happy to share my contribution to "Parcours IA en Santé", the first 🇫🇷 podcast on integrating
#AI
in healthcare.
🧠 We delve into my optimistic take on automated machine learning (
#AutoML
), explainability, cherry-picking 🍒, etc...
🔗 Listen ➡️ (FR)
@n8radcliffe
I think it’s fixation light, but I was never able to make it work. I suggest you get rid of it and keep asking patients to look at your ears 😂
Art and the
#eye
.
Check out our piece in JFO 🇫🇷. We explore the anatomy of a rose 🌹 to describe a case of Valsalva retinopathy.
We then revisit Saint-Exupéry's Little Prince 🤴🏼 and discuss what he learns from his own rose.
#Ophthalmology
#Retina
#Art
Happy to share my interview with
@mlabrosse
on
@le1518
🎙️(in french)! Explored: my journey from Syria to Quebec, carrying forward my father’s
#retina
legacy, the potential of
#AI
, Syria’s refugees & how my charity Fondation d'Alep is responding.
[3/4] We are happy that this was published in the Saudi Journal of Ophthalmology 🇸🇦, which is widely read in the Middle East and North Africa. Read the article here:
@DrXiaoLiu
Hey
@DrXiaoLiu
! I noticed that too. I’ve actually trained an AutoML model using one of those datasets. The saliency maps seem to highlight the vertebral bodies in the children’s CXRs.
@kevinelaimd
Read 25% of the book pre-rotation, and finished by mid-rotation. Used to create a schedule. Printed it, glued it to the book cover, and highlighted progress.
We are excited about the potential of ChatGPT in ophthalmology but remain very cautious. Imminent clinical applications seem limited. How will LLMs handle imaging data like OCTs? [4/4]
ChatGPT did best in general medicine and worst in neuro-ophthalmology and ophthalmic pathology/ intraocular tumors. This is interesting because those are highly-specialized domains that are considered challenging even within the ophthalmology community. [2/4]
@n8radcliffe
What advantage do you see of endoscopy in MIGS? 😊 Poor image quality is a limiting factor. I prefer heads-up displays like NGENUITY for better ergonomics, high magnification and really an amazing view of the angle!
Check out our new photo essay in
@CanJOphth
: the 1st published case of
#brolucizumab
-associated inflammation in Canada. Consent should discuss the risk of irreversible vision loss & looking for
#uveitis
before every injection is crucial.
#ophthalmology
4/ More work is needed, but exciting times are on the horizon as VR and metaverse technologies continue to advance.🛠️ RetinaVR is a collaboration between
@chumontreal
&
@polymtl
. Funded by the 2021 Innovation in Retina Research Award of the
@CANeyeMDs
&
@Bayer
.
[2/4] We then integrated the best model into an Android application that utilizes Edge architecture, enabling offline diagnoses without recurring costs.
We wrote a comment in
@Eye_Journal
regarding 360° laser retinopexy for the prevention of primary surgical failure after RRD repair. We discuss future directions and the study design for prospective superiority trials.
#Ophthalmology
#retina
[1/4] Without writing a single line of
#code
, our clinicians created powerful deep-learning models to diagnose trachoma from conjunctival images. This is the power of AutoML.
@chumontreal
@polymtl
@CANeyeMDs
@Bayer
5/ Note: This video is a demonstration of our app during pilot testing on the Meta Quest 2. Work is currently undergoing peer review. Refer to our preprint for limitations. RetinaVR is not yet available for download. ends/
Docs are ROCs: A simple fix for a methodologically indefensible practice in medical AI studies.
Widely used methods to compare doctors to
#AI
models systematically underestimate doctors, making the AI look better than it is! We propose a solution.
1/7
🎯 Exam section & question difficulty were most predictive of ChatGPT's answer accuracy. Legacy model: best in general medicine, worst in neuro-ophthalmology & ocular pathology. ChatGPT Plus: no differences between subspecialties = improved consistency!
#pathology
#neurology
At tonight's
@UMontreal
retina journal club, I'm presenting this great study by
@hbogunovic
and team, published in
@AAOjournal
.
#AI
-based quantification of fluid in AMD may soon replace CRT in AMD trials!
ChatGPT was more accurate when a higher percentage of human peers obtained the right answer for a specific question. This is comforting as it suggests that ChatGPT’s responses align, to a certain degree, with the collective understanding of ophthalmology trainees. [3/4]
2/ RetinaVR is an app built for commercially-available VR headsets and their standard controllers.🎮 It leverages their powerful sensors and cameras, without the need for expensive external haptic devices. This dramatically reduces cost.
🚀 In conclusion, ChatGPT - an evolving publicly available LLM - shows potential in ophthalmology Q&A. Do we need domain-specific LLMs (e.g. trained on EyeWiki) to get better performances, or is GPT-4 the answer? 🤔💡
#GPT4
3/ We focused on 4 fundamental skills: core vitrectomy, peripheral shaving, membrane peeling and endolaser application. We ran a pilot validation study of 20 participants to establish construct validity and to gather feedback. 💬
@iDrFeng
Hey
@iDrFeng
, couldn't agree more! Exciting times indeed. While
#ChatGPT
takes the wheel for the questions, it's worth mentioning - the brilliance behind this tweet might just be the AI's doing. How's that for a sprinkle of AI magic on your podcast? 🤖🎙️
ChatGPT achieved 55.8% and 42.7% accuracy in the two 260-question simulated exams using the BCSC Self-Assessment Program (
@aao_ophth
) and OphthoQuestions. [1/4]
Happy to have won the best resident paper award at the
@UMontreal
Ophthalmology Research Day. We developed a model using AutoML to predict PVR formation after retinal detachment repair.
#ophthalmology
#machinelearning
📊 Legacy ChatGPT scored 55.8% on BCSC set & 42.7% on OphthoQuestions set. ChatGPT Plus improved to 59.4% & 49.2% respectively. We noted that easier questions = higher accuracy! 🧠
#LanguageModel
@3Y3MD
@ArevaloJF
Thanks for sharing
@3Y3MD
, and congrats
@ArevaloJF
! This inspires me to continue my father's legacy, just as he has. I met Dr. Arevalo at the 2017 CRS meeting in Whistler🇨🇦 as a medical student, and I fondly remember his expertise & kindness as he sat on the panel during my talk.
@DrewCareyMD
They may perform better. The real breakthrough, I think, will be the use of multimodal models (like CLIP) that combine knowledge of text and images.
@kevinelaimd
Of course! Building on your slide on Bloom's taxonomy, I will add: I did the BCSC and the OQ questions every single year. That's ~ 9K q/yr x 4 yrs = 36K questions = only 25q/day! Consistency is key. Recommend creating tests with all categories combined for spaced repetition.
@surgeonretina
Racemose angioma (retinal AVM) can be isolated or associated with Wyburn-Mason syndrome. The syndrome presents with multiple AVMs in different locations (brain, retina, orbit, etc). A brain MRI can help distinguish the two entities (looking for a midbrain AVM).
🚨 Lack of image processing (for now) limits usefulness. Accurate differentiation between superior RD (needing PnR) & inferior RD (requiring SB), for example, might remain a challenge due to insufficient domain-specific image-text pairs in training data.