Introducing Hibou, a family of state-of-the-art large foundation models in pathology from HistAI

HistAI is proud to announce the launch of its inaugural family of Vision foundation models, Hibou-L and Hibou-B. These models have undergone extensive training on over 1.1 million whole slide images, leveraging one of the most diverse proprietary datasets ever assembled for AI model development in pathology. The training data encompasses a comprehensive range of human tissues and organs stained with routine H&E, as well as hundreds of other stains. Furthermore, the dataset is enriched with veterinary and cytology cases, enhancing the models' generalizability and applicability to a wide spectrum of pathology studies.

HistAI - AI-native challenger in computational oncology announced the release of its first family of Large Vision Foundation models to accelerate the pace and adoption of highly accurate and performant AI-algorithms pushing the limits of drug development pipelines as well as laboratory diagnostics.

In the ever-evolving landscape of machine learning, foundation models stand out as groundbreaking innovations. These models undergo self-supervised learning, processing vast amounts of unlabeled data to uncover and encode critical patterns and information. Once trained, foundation models become incredibly versatile, capable of adapting to a wide array of specific tasks and new, unforeseen contexts.

"I am incredibly pleased to announce the culmination of four months of hard work by our entire team— the release of our first foundation models. I am especially proud to uphold our commitment to open access to AI innovations by making our models accessible to as many researchers and developers as possible, either for free or at a nominal cost that pales in comparison to the expenses associated with datasets and training. I believe that the release of our top-performing foundation model under the Apache 2.0 license will significantly accelerate innovation in computational pathology, leading to tangible benefits for humanity as a whole." said Alex Pchelnikov, CEO an co-founder of HistAI.

Utilizing high-quality Vision Foundation model in pathology significantly enhances the development of downstream algorithms, offering greater efficiency and cost-effectiveness compared to creating custom tools from scratch. For instance, a vision foundation model in pathology can seamlessly transition between tasks such as tumor classification, region of interest (ROI) segmentation, cell segmentation, tumor subtyping or grading.

"I am excited to be part of this pivotal moment in computational pathology. The introduction of our Vision foundation models represents a significant leap forward in harnessing AI for cancer research. By offering these state-of-the-art tools with open access, we are enabling the research community to drive innovation and achieve groundbreaking advances in oncology. Our goal is to push the boundaries of what is possible in cancer diagnostics and drug development, ultimately transforming patient care." said Katherine Ivanova, COO and co-founder of HistAI.

Hibou-L, HistAI's flagship model, has been meticulously trained on 1.2 billion pathology slides. It has established itself as the state-of-the-art vision foundation model in pathology by achieving the highest scores across nine common public benchmarks. Starting today, it is available to "Scale" subscription users of the HistAI CELLDX platform via API.

Hibou-B is freely distributed under the most permissive open-source license (allows commercial use) and is available for download on the Hugging Face platform.

HistAI is open to collaborative research and is delighted to offer preferential access to its Vision Foundation Models to academic institutions worldwide. All requests concerning access to HistAI's models can be sent to models@hist.ai or via contact form on www.hist.ai