According to Andrew Ng (@AndrewYNg), a U.S. researcher specializing in the field of Artificial Intelligence, AI is not yet used in hospitals around the world for the following reasons: the scarcity and poverty of data in the field of medicine especially for rare diseases, the difficulty of researchers to deploy their models on real cases, the lack of change management induced by AI in the health sector.
Dealing with the scarcity of labeled data
2 billion chest X-rays are performed annually to detect pneumonia, cancer, Covid-19. Supervised deep learning has reached the level of radiologists in 11 out of 14 pathologies.
Supervised learning is an automatic learning task consisting in learning a prediction function from annotated examples. For the detection of pleural effusion, AI does as well as the radiologist, but this is not the case for the diagnosis of thoracic hernia.
For Pleural Effusion, the annotated data set is 11000 examples compared to 110 examples for hernia. The number of correctly annotated data is one of the determinants of the effectiveness of supervised learning. In medicine for certain pathologies, these annotated data do not exist or are too few.
To overcome the scarcity of data, Andrew Ng suggests several methods already used by researchers: transfer learning, one-shot learning, self-supervised learning, data augmentation.
Transfer learning
The situation in which what has been learned in one context is exploited to improve generalization in another context. Transfer learning should allow knowledge gained from previous tasks to be used and applied to more recent tasks.
One-Shot learning
While most object categorization algorithms based on machine learning require training on hundreds or thousands of samples or images and very large datasets, ‘learning at once’ is about learning information about object categories from a single or a few samples.
Self-supervised Learning
The idea behind ‘self-directed learning’ is to develop a deep learning system that can learn to fill in the gaps. “You show an input item, a text, a video, or even an image, you remove part of it, you hide it, and you train a neural network to predict the missing part. This can be the continuation of a video or missing words in a text.
Data Augmentation for Deep Learning
There are several image enhancement techniques (perspective transformations, contrast changes, Gaussian noise, hue/saturation changes, cropping, blurring) that can be applied to the image to enhance learning.
Generalize models with new data
Too many research papers propose models that cannot be used in production. Radiologists do not have the same equipment throughout the world, the quality of the images is not necessarily the same as that of the images used to train the models. To overcome this recurring problem, Andrew Ng advises putting the technicians, those who know the radiography equipment, in the loop. They can inform the researchers about possible technical biases as soon as the model is created and trained.
“Like most image recognition systems, the deep learning model was trained on high-quality scans; to ensure accuracy, it was designed to reject images that fell below a certain quality threshold. Because nurses in Thailand scan dozens of patients per hour and often take pictures in poor lighting conditions, more than one-fifth of the images were rejected.”
Change Management
Finally, Andrew Ng points out that the real challenge lies in the perception of Artificial Intelligence by the health professions. How to engage healthcare workers to use AI in the interest of patients without feeling dispossessed of their job and responsibilities.
AI will define new processes of care and it will be necessary to involve everyone in the interest of patients. Caregivers need to understand that the computer, the robot, the algorithm will not replace them but will help them to be more efficient and have more time to spend with patients.
Doctors Are Confident That AI Won’t Replace Them-Forbes-Janvier 2019
“Researchers highlight a gap between the medical community and the AI community. With the gradual introduction of AI into health systems around the world, they fear that the medical community’s skepticism about the validity of the technology may lead to an inevitable conflict.”
INDEX
Article written from this video : Pie & AI: Real-world AI Applications in Medicine. (1:10' à 1:23': Andrew Ng)
