AI-centric Approach, Exceptional Technology
Lunit’s research team is committed to addressing the fundamental challenges of AI.
We define general AI problems from domain-specific issues and solve it in our own way.
We target 99% accuracy, always
We believe that the medical field is a special area that requires an utmost level of accuracy. Even a difference by 1% in accuracy can make life-saving changes.
Currently our Chest Radiography and Mammography solution shows 97-99% and 97% accuracy, respectively. Lunit’s goal is to always achieve 99%.
And we use our unique, state-of-the-art AI training technology to achieve unprecedented accuracy.
We use “Human in the Loop” simulation
Annotation is important in training AI with medical data. Traditionally, physicians pick data for annotation, after which, the annotated dataset is used for training. The problem is that many of these annotations often contain redundant or trivial examples, which is not the best particular method in improving the performance of AI.
Our AI actively suggests the cases that it perceived as “most uncertain” within each given dataset. These cases are then provided with confirmed annotations from our physicians. Think of it as, “learning from mistakes.”
Making the most of the acquired data
Data is flooding in, but AI is still data-hungry. Finding the most effective way to annotate a vast amount of medical data is one thing - making use of all the rest is another. We want our AI to learn from all the data, but the budget for annotations is limited.
To solve this issue, we employ unsupervised methods that exploit what the AI thinks it knows about the unlabeled data. After all, our AI learns by itself with unlabeled data, and learns by asking the physicians for uncertain data.
Our AI share knowledge between tasks
Once you speak French, it is easier to learn Spanish.
We believe that the various tasks within medical image analysis share primitive knowledge.
By explicitly guiding the AI training process to find and share the common knowledge in between different areas of expertise, we can equip our AI with more robust and scalable knowledge.