Research

The Orthopedic Surgery Artificial Intelligence Laboratory (OSAIL) has several key research focus areas.

Image annotation

Manual annotation of medical images is time-consuming, error-prone and subject to interreader variability. Our lab develops cutting-edge deep-learning algorithms that automate these tasks. This allows for accelerated creation of labeled datasets and algorithms that benefit from an abundance of data.

Risk stratification

Joint failure after joint replacement surgery is rare but does happen, and it can be catastrophic for the person undergoing the procedure.

OSAIL is a leading advocate of personalized medicine in orthopedics, where individual characteristics are used as inputs for models that predict the risk of postoperative complications like joint failure. These models help surgeons optimize their approach for the best postoperative outcomes that mitigate risk, reduce cost, improve quality and enhance patient satisfaction.

Imaging registries

A common limitation in data science is that most of a researcher's time is spent organizing and cleaning data for further analysis.

Our lab has invested heavily in the creation of large-scale imaging registries with images pulled from Mayo Clinic's electronic health record system and automatically labeled for future use with bespoke annotation algorithms. These registries enable rapid project progression and large cohort studies that would otherwise be prohibitively laborious.

Text processing

Using large language models, OSAIL develops natural language processing (NLP) models that can parse and synthesize text data, including operative notes, progress notes and lab values.

Uses for these NLP models include distilling unstructured data in conjunction with our imaging registries and predicting surgical outcomes from automatically pulled patient characteristics.

Synthetic data

Image generation is the umbrella term for deep-learning projects that aim to generate synthetic imaging data resembling real imaging. Synthetic images protect patient privacy and allow for substantially larger datasets for algorithm development.

Our lab has created algorithms that generate synthetic imaging datasets, improve the quality of existing images through resolution upscaling or artifact reduction, and modify real images with synthetic pixels.