Deep Learning is a great AI technology that we read about everywhere. Modern cloud computing and graphics processing units allow this technology from the 1980’s to be used efficiently today. It has been a leap forward for many types of AI algorithms to use these Artificial Neural Networks to successfully interpret medical images. Deep learning developers need to consider clinical rigour for medical devices. So why is this important and what do we need to look out for?
Training AI Algorithms
Deep Learning accuracy relies on a quality training data set. The training data set trains the deep learning algorithm. Once trained you can adjust the settings of your algorithm to improve your results. Normally, artificial neural networks learn continuously. Once trained, medical AI algorithms are locked. This is a regulatory requirement to prevent skewed data in the clinical environment changing the accuracy of the algorithm.
Every aspect of your training data effects the accuracy of your medical imaging algorithm. The variables in the data that can affect accuracy include:
- image quality
- make and model of modalities
- patient ethnicity, age and gender
- pathology types detected
- anatomical structures
- implantable devices/prosthetics
The only solution is for manufacturers to keep adding new training data to ensure the right mix of all the possible variations are in the data.
Clinical rigour is when AI vendors take this algorithm training seriously and continually check and recheck the validity of their training model. Every time they take their product to a new country – they check and recalibrate for the local ethnic mix. In the case of medical imaging AI algorithms, even the make and model of imaging modalities is an important factor. High quality AI vendors train their AI algorithms using hundreds of thousands to millions of studies. They take all the data variations listed above into account.
An AI product that has been rushed to market can have its accuracy drift from the published results when used with untested data variables in the clinical environment. This is why clinical rigour is important. A medical AI solution is of no use if it has 95% accuracy in the lab but only performs at 60% in the real world.
Kheiron Medical’s Mia AI Software for breast screening is a great example of an AI product developed with deep clinical rigour.
AI Advance is proud to be the distributor of Mia in ANZ.