Case Study: Harvard Medical School Researchers build a 3D CT foundation model with Lightly

A Lightly Case Study

Preview of the Harvard Medical School Researchers Case Study

Harvard Medical School Researchers build a 3D CT foundation model with Lightly

Harvard Medical School researchers, part of the AI in Medicine Program at Mass General Brigham, faced significant challenges in training a general-purpose 3D CT segmentation model. Their medical imaging pipelines were highly heterogeneous, and existing self-supervised learning (SSL) repositories were not designed for 3D data or CT-appropriate augmentations. They needed a clean, modular SSL workflow to unify experiment setups across their lab and move beyond task-specific tuning.

Lightly provided a solution with its LightlySSL product, delivering a clear and well-structured DINOv2 implementation that was easily adapted for 3D CT volumes. The solution integrated seamlessly with the lab's existing MONAI and PyTorch Lightning setup, enabling CT-specific augmentations and a reproducible training environment. This allowed the researchers to build a foundation model that improved feature-level representations, served as a stable baseline for multiple studies, and accelerated internal experimentation by reducing engineering complexity.


View this case study…

Lightly

10 Case Studies