Mohammed Baharoon

سعيد جدًا أنني سأنضم إلى برنامج الدكتوراه في الذكاء الاصطناعي في الطب ضمن تخصص المعلوماتية الطبية الحيوية في جامعة هارفارد، ابتداءً من خريف 2026، بدعم كامل من الجامعة. ممتن لجميع من دعمني ووجّهني في هذه الرحلة. شكر خاص لمشرفي الحالي الدكتور Pranav Rajpurkar، ولمشرفي الأول وصديقي الدكتور عبدالرحمن الجوعي، وكذلك لمشرفيّ خلال فترات التدريب السابقة الدكتور Bo Wang و الدكتور Dominik L. Michels.

Extremely excited to share that I will be joining the AI in Medicine Ph.D. program (@AIM_Harvard_PhD) in Biomedical Informatics at Harvard Medical School/Harvard Kenneth C. Griffin Graduate School of Arts and Sciences (@harvardmed/@HarvardGSAS). Grateful to all my mentors for their guidance along the way. I am especially thankful to my current advisor Prof. Pranav Rajpurkar (@pranavrajpurkar), my first advisor and friend Prof. Abdulrhman Aljouie (@aaljouie), and my mentors from previous internships and studies, Prof. Bo Wang (@BoWang87) and Prof. Dominik L. Michels (@dominikmichels).

Thank you very much to our spotlight demo presenters! @VasishtSumukh @BaharoonMS #ML4H2025

@BoWang87 @UHN @UofT @VectorInst Congratulations!!!

Honoured to be officially inducted today into the Royal Society of Canada’s College of New Scholars. 🇨🇦✨ Grateful for my students, collaborators, and institutions who made this journey possible. @UHN @UofT @VectorInst AI, genomics, and precision medicine are converging faster than ever—and I’m excited to keep pushing the frontier, from SNF, scGPT, MedSAM to BioReason to the emerging vision of the virtual cell. @Xaira_Thera Onward. 🙏🧬🤖

Thrilled to share that our paper, “Exploring the Design Space of 3D MLLMs for CT Report Generation,” has been accepted to @MICCAI_Society 2025! 🎉 This work—developed during Mohammed Baharoon’s internship at the Vector Institute with Jun Ma (@junma)—secured 2nd place in the MICCAI 2024 AMOS-MM Challenge 🥈. We systematically explored how to build effective 3D multimodal large language models (MLLMs) for CT report generation, focusing on: --Architectures: visual encoders, projectors, and language models --Training strategies: from frozen LLMs to parameter-efficient fine-tuning --Report completeness: knowledge-driven methods to enrich outputs 💡 Key innovations: --Binary-based Questioning (BQ): prompts the model to ask yes/no questions about common findings, reducing missed abnormalities. --Naive Normality (NN): automatically inserts “normal” statements for unmentioned organs, improving clarity and completeness. Together, these boosted the GREEN score by up to 10%, showing how simple post-processing can make AI-generated reports more clinically accurate. Takeaways: 1. Architecture matters more than LLM size 2. Higher resolution isn’t always better if pre-training and fine-tuning resolutions mismatch 📄 Full paper: papers.miccai.org/miccai-2025/pa… 💻 Code: github.com/bowang-lab/AMO… Shoutout to Mohammed Baharoon and Jun Ma for their leadership in this project! Also thanks to all other co-authors, Congyu Fang and Augustin Toma, for their tremendous contributions! 🙏 @VectorInst @UHN_Research @UHNAIHUB @UofTCompSci

It's the 17th anniversary of Castle Crashers' initial release on the Xbox Live Arcade! Game babies, they grow up so fast. 💗⚔️

Excited to share #Harmony: A joint self-supervised and weakly-supervised method learning generalized visual features from web-scraped data. It outperforms #CLIP/#SigLIP/MaskCLIP/SLIP/iBOT/MAE on classification/segmentation/detection. arxiv.org/abs/2405.14239 @KAUST_News @Harvard

@gabriberton Used a similar technique here thanks to ur original tweet:) arxiv.org/abs/2405.14239 [2405.14239] Harmony: A Joint Self-Supervised and Weakly-Supervised Framework for Learning General Purpose Visual Representations

The TIPS paper by Google is a perfect example of a method where using my PyTorch trick would greatly reduce GPU memory need Note how the 3 losses are disentangled In this case you can cut memory by 3x (!!!) if you backward separately on the 3 losses, and get identical results

@qw6fr مبروك:)

الحمدلله ! فزت بقيادة نادي قوقل للسنة القادمة. 🥳 #KSAUcore

@iMohannadG تخسي هوليوود

#خريجو_كاساو_2025 مبارك تخرجكم 👏🏼 والله يوفقكم ويبارك دربكم دومًا للأفضل 🤍

MedSAM2: Segment Anything in 3D Medical Images and Videos "fine-tuning the Segment Anything Model 2 on a large medical dataset with over 455,000 3D image-mask pairs and 76,000 frames, outperforming previous models across a wide range of organs, lesions, and imaging modalities. "

Ready for the most powerful foundation model for medical images/videos? 🚨 Just dropped: MedSAM2 The next-gen foundation model for 3D medical image & video segmentation — built on top of SAM 2.1. • Project site: medsam2.github.io • Paper: arxiv.org/abs/2504.03600 Why it matters: • Trained on 455K+ 3D image–mask pairs & 76K+ annotated video frames • >85% reduction in human annotation costs (validated in 3 studies) • Fast, accurate, and generalizes across organs, modalities, and pathologies Big impact: We used MedSAM2 to create 3 massive datasets: • 5,000 CT lesions • 3,984 liver MRI lesions • 251,550 echo video frames Plug & play: Deployable in: → 3D Slicer → JupyterLab → Gradio → Google Colab We open-sourced everything! Explore more: • Models: huggingface.co/wanglab/MedSAM2 • Datasets:  → CT: huggingface.co/datasets/wangl…  → MRI: huggingface.co/datasets/wangl… • 3D Slicer plugin: github.com/bowang-lab/Med… • Segmentation dataset hub: medsam-datasetlist.github.io Let’s collaborate: Have unlabeled 3D medical images or videos? We’re actively seeking partners to co-create new public datasets. DM us or connect! #MedSAM2 #MedicalAI #ComputerVision #3Dsegmentation #FoundationModels #AIforHealth Also shoutout to @JunMa_11, the machine learning lead at our @UHNAIHUB for his amazing leadership in this project!

I'm happy to announce that I can't take it anymore

@iScienceLuvr ;)

Thinking about creating a medical AI group chat here on Twitter... If you're a researcher/engineer/clinician/etc. working in AI or medical AI, let me know if you're interested in joining!

@giffmana !!

Interesting

لهنت لهنت

New junior developers can't actually code. Fantastic blog where the author argues that AI is preventing devs from understanding anything Unlike the internet which only replaced fact lookups but adds explanations, AI replaces all reasoning.

True. SimCLR and BYOL are technically JEA, not JEPA. But the idea of contrastive SSL for JEA is much older (1993L with updates in 2005 and 2006. - Signature Verification using a "Siamese" Time Delay Neural Network. Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, Roopak Shah. NIPS 1993 - Learning a similarity metric discriminatively, with application to face verification. Sumit Chopra, Raia Hadsell, Yann LeCun, CVPR 2005 - Dimensionality reduction by learning an invariant mapping Raia Hadsell, Sumit Chopra, Yann LeCun, CVPR 2006. There was connected work in metric learning by David Lowe, Killian Weinberger, and Salakhutdinov/Roweis/Hinton