Presenting the entering class of 2025, UCSF Radiology

Being able to work on and finish an RCT during residency has been a dream come true! Wouldn’t have been possible without our amazing team Jae Ho Sohn, MD, MS UCSF Cardiothoracic Imaging UCSF Imaging UCSF IR 🥰 please look out for our manuscript!

In this editorial, Drs. Jae Ho Sohn, MD, MS and Brandon K.K. Fields, M.D. discuss the study by Pan and Hu et al for predicting preinvasive, minimally invasive, and invasive adenocarcinoma. UCSF Radiology Residents UCSF Imaging UCSF Center for Intelligent Imaging UC San Francisco @UCSFchest doi.org/10.1148/radiol…

3⃣ Victor Quach, Adam Fisch, Tal Schuster, YalaTweets, Jae Ho Sohn, MD, MS, along w/ #JameelClinic PIs Tommi Jaakkola & Regina Barzilay propose an extension of conformal prediction tailored specifically to generative language models. Check it out on 5/8 in Session 3: arxiv.org/abs/2306.10193

UCSF medical student and 0.55T lung MRI researcher, Felicia Tang, commenting on her recent research and potential future directions for lung MR: auntminnie.com/resources/conf…

Extremely excited to join Jae Ho Sohn, MD, MS's lab UCSF Center for Intelligent Imaging as a T32 postdoctoral research fellow! Really looking forward to all of the great #radiology #ai research that we are going to do!

Excited to moderate #RadAIchat on July 3rd, 8 PM ET! We'll explore BERT models in radiology with panelists Chunnan Hsu PhD Jae Ho Sohn, MD, MS Daniel Truhn Keno Bressem Join us to discuss how these AI models are shaping medical imaging. Interesting questions ahead! Hope to see you there.

T1. BERT is a type of large language model that takes into account context of words in a sentence when doing natural language processing. It was a significant leap and initial step breakthrough towards the current large language models that we see in the field. #RadAIChat

#RadAIChat T2. Practically speaking from a user perspective, the key difference of BERT and GPT is that BERT is usually fine tuned towards a specific task whereas GPT is designed for generation. GPT, when large enough, typically performs well even without finetuning.

T4. Similar to other AI clinical translation challenges - Generalizability, fairness, bias related issues; AI governance & regulatory approval; post-surveillance monitoring; real-time access to data if real-time integration is desired #RadAIChat

T5. I'd highly recommend our RSNA Deep Learning Lab's NLP lectures. They introduce you to NLP (including BERT) that would be practically relevant for radiology. Course Github from 2023: github.com/RSNA/AI-Deep-L… #RadAIChat

Some amazing lab members and collaborators visiting this summer! UCSF Imaging UCSF Center for Intelligent Imaging

LAPD raid goes from bad to farce after gun allegedly sucked onto MRI machine sfgate.com/cannabis/artic… via SFGATE

It was wonderful catching up with everyone at #KCR2024! Grateful for the opportunity and a special thanks to Dr. Changmin Park and #KSTR for the warm hospitality. SNU Radiology UCSF Imaging

The 8 – 12 week summer internship program at @UCSF_ci2 is open to all undergrad, grad & medical students. Don't miss this chance to collaborate with researchers on #AI & #MachineLearning projects! Repost to share with your networks! bit.ly/4jhdLaq

Can smarter search boost diagnostic AI? 🧠 A study by Cody Savage, Gunvant Chaudhari, Andrew Smith & Jae Ho Sohn, MD, MS shows that RadSearch, a specialized semantic search model, improves radiology report retrieval and boosts LLM diagnostic accuracy. See how it works! pubs.rsna.org/doi/10.1148/ra…

Farewell dinner to Tician. Thank you for the wonderful 1.5 years here at UCSF. Looking forward to seeing you in your next steps in your career back in Switzerland. Looking forward to all of our future collaborations!!

It's #NationalInternDay & we're celebrating our group of summer interns at UCSF Center for Intelligent Imaging. Visit bit.ly/4l2niC1 to read more about Koharu Sakiyama & her work with Dr. Jae Ho Sohn.