I am excited to share our #CVPR2024 DEF-MIA paper on the Interactive Medical Image Learning Framework (IMIL) 🩻!

We showed that IMIL could effectively use clinician feedback on just 4% of the dataset to yield a 4.2% improvement in accuracy! 🚀

🧵 below
arxiv.org/abs/2404.10965

Should we stop training software engineers Geoffrey Hinton ? 😉

Kathi_Hoebel Caroline Chung, MD Ken Chang Jayashree Kalpathy study experts' ratings of brain tumor segmentation doi.org/10.1148/ryai.2… QTIM Lab MGH Martinos Center #DeepLearning #AI #ML

After an year long tussle obtaining a J1 waiver, a visa rejection, reaching out to my local Senator & recovering from Covid, finally starting on a new journey as Assistant Professor at CU School of Medicine.

Thanks to all my mentors & well wishers for making this possible ❤️🥳✊

Congratulations to the newly minted Twitter-less Jay Patel, Ph.D.!! What a ML rock star! Big ups to his mentorship committee Jayashree Kalpathy, Elizabeth Gerstner, Bruce Rosen, and Elfar Adalsteinsson! MGH Martinos Center HST

Collaborative training of deep learning models in medical imaging is key to improving their robustness and generalizability. Our latest article reviews methods for collaborative training and key considerations for successful deployment. RadioGraphics pubs.rsna.org/doi/10.1148/rg…
1/n

I’m #PhDone ! Last week I defended my PhD thesis “Domain and User-centered ML for Medical Image Analysis. Big thanks to my fantastic advisor Jayashree Kalpathy and my committee 𝗣𝗼𝗹𝗶𝗻𝗮 𝗚𝗼𝗹𝗹𝗮𝗻𝗱, 𝗕𝗿𝘂𝗰𝗲 𝗙𝗶𝘀𝗰𝗵𝗹, and 𝖉𝖆𝖛𝖊 𝖋𝖚𝖑𝖑𝖊𝖗 @[email protected]. Now on to revisions & papers!

slow clap for the newly-minted Kathi Hoebel, M.D., Ph.D. 🧠🎉 we will miss you! Kathi_Hoebel MGH Martinos Center Jayashree Kalpathy

This is an AMAZING paper!

Basically showing how *mistakes* in data handling and model development can make an ML model with actual AUC of 0.55 have and apparent AUC of 1.0.

The authors do lots of experiments to pinpoint some of the most important issues, fantastic!👌🏻

T1 Saliency maps are used to mitigate the “black-box” nature of AI algorithms. Arun et al. highlight a major limitation of the saliency map technique, including limited repeatability/reproducibility on localization, segmentation and detection tasks in medical imaging. #RadAIChat

To avoid overestimation of the causal relationship between radiomic features and other variables, computational experiments revealed that the data set size should equal or exceed the number of radiomic features under consideration.​ bit.ly/3WqeBWk

Great to see this out!
ACR’s Connect and AI-LAB technical framework academic.oup.com/jamiaopen/arti…

Just met w my American College of Radiology mentee Dr. Matt Li @ #RSNA22 for the 1st time in person!

He’s now joined University of Alberta University of Alberta_FoMD in #Edmonton - adding to the amazing 🇨🇦 #radiology community

Huge 🙏 to the ACR for asking me to be a mentor. The best outcome I could have asked for!

The Boston branch of QTIM is seeking stellar students and programmers interested in #deeplearning and medical imaging. Please share or reach out ([email protected]) for more info!

.Eliot Siegel has frequently promised that he would wash the car of anyone who built an algorithm that could segment the adrenals on CT better than a 5th grader.

Time to get the chamois ready...
pubs.rsna.org/doi/10.1148/ra…

A big congrats to our wonderful lab co-director, Dr. Elizabeth Gerstner, on winning the @AANMember Neuro-oncology Scientific Award!

Loved working with fantastic HST students Nalini Singh Jordan Harrod Sandya Subramanian Ken Chang Mitchell Robinson & our all star faculty. Hope this inspires more great work!

More than a year after the symposium, the paper is finally out! doi.org/10.1007/s12021…
Thank you to Randy Gollub for inviting me to be a part of this effort!
Nalini Singh Jordan Harrod Mitchell Robinson Ken Chang