LG AI Research Open-Sources EXAONEPath, a Model Specialized for Histopathology Images

LG AI Research has open-sourced EXAONEPath, a model specialized for processing histopathology images. Unlike general images, histopathology images have unique characteristics: they are extremely large (20k x 20k or more), have a distinct range of color tones, and contain limited objects.

EXAONEPath achieves accuracy similar to competing models from global big tech companies, while reducing the model size to about 1/10 the level, securing overwhelming cost efficiency. Its key feature is the ability to deliver excellent performance even with a small amount of training data and low infrastructure costs.

With this model enabling AI to analyze histopathology images, it becomes possible to predict genetic mutations or determine suitable treatments while skipping the genetic testing step, which previously took up to 2 weeks, dramatically reducing time and cost.

The key technical points are as follows:

• Utilizes the MIL (Multi-Instance Learning) framework to process gigapixel-sized images on a patch-by-patch basis.
• Applies the DINO approach among self-supervised learning techniques to secure high-quality representations even from unlabeled data.
• Maintains scale consistency by incorporating meta information so that all patches have the same MPP (Micro-meter Per Pixel).