LG AI Research 470

LG AI Research has released EXAONEPath, a pre-trained model optimized for the analysis of histopathological images, as open source. Unlike general images, pathology images are extremely large (approximately 20k x 20k) and have unique color ranges and objects (nuclei, cytoplasm, etc.), making a specialized model essential.

EXAONEPath is a model that achieves both high performance and cost-effectiveness. Across 6 benchmark results, it showed accuracy on par with global competitor models, while demonstrating high efficiency with a size of only 1/10 compared to competitor models and requiring less training data. This means infrastructure costs can be significantly reduced.

This model aims to drive innovation in the bio field. Once AI can analyze pathology images, it will be able to predict genetic mutations and determine appropriate treatments and drugs without separate genetic testing. This can shorten the genetic testing period, which previously took up to 2 weeks, saving both time and cost.

The core of pathology image research is WSI (Whole Slide Image) processing. General AI models are trained by downscaling images to 256~1024px, but downscaling WSI makes it impossible to identify the shape of cells. Therefore, the MIL (Multi-Instance Learning) framework is widely used to address this issue.