I totally agree about writing being a tool for thinking - but LLMs have had the opposite effect on me.

I write out my thoughts about something to them. In the process of describing the task at hand, I think about it more deeply and get much better insight.

@eugenevinitsky.bsky.social

Weird! I see myself in it. But I also see you. Let me try removing and adding you again.

In the paper we highlight additional interpolation abilities - like the ability to close a specific cartilage hole.

Left: recon of patient data.
Moving right, we slowly close the cart hole...

This capacity opens the door for new types of in silico simulation!

Our hybrid implicit-explicit Neural Shape Model outperformed all others in reconstruction tasks and most prediction tasks. 🏆

This highlights the power of deep learning-based shape modeling, surpassing both traditional shape models and CNNs in performance.

Prediction Benchmarks:
Our benchmarks include predicting OA grade and more advanced metrics like the MRI Osteoarthritis Knee Score (MOAKS).

MOAKS enabled evaluation of whether a model can localize specific features of disease like osteophytes, cartilage thinning, and cartilage holes!

Recon Benchmarks:
We include traditional surface error reconstruction metrics like ASSD. We also compute cartilage thickness biomarkers to determine how well our models preserve these data that are currently used in research and clinical trials.

Benchmarks:
Datasets exist for shape modeling, like ShapeNet and even recently MedShapeNet. However, they lack clinically relevant evaluations to assess if the model captures critical medical info.

We created benchmarks that evaluate both traditional recon and clinically relevant measures...

I'm excited to share that our work on #deeplearning based shape modeling is finally out in #ieee.

🧵 for info about Data, Benchmarks, and Models!

@akshay-chaudhari.bsky.social @stanfordmedicine.bsky.social

www.medrxiv.org/content/10.1...
ieeexplore.ieee.org/document/107...