AI’s increasing role in medical imaging
Medical imaging is a highly effective tool for diagnosing a wide array of diseases and injuries, but it often requires expert-level skills to interpret accurately. AI techniques such as deep learning have revolutionized image analysis, and are expanding the reach and improving the quality of medical imaging.
Companies like AIDoc have begun providing computer-aided diagnostics that integrate into radiologists’ existing workflows. Meanwhile others, like Behold.ai, are driving down the cost of radiology services. Rather than replacing human analysis, radiologists are turning to these AI-based systems to improve their diagnoses.
Radiologists face a common problem affecting many medical diagnosticians: Data about patients is available in greater volumes and in more complex forms. Healthcare providers may find that they have more cases in which imaging could be useful but are unable to employ it because skilled radiologists, who are needed to interpret the results, are unavailable or already working at capacity.
New advances in image processing are well suited to medical applications, a domain that has widely adopted informatics tools; in fact, medical practitioners were early adopters of decision support systems dating back to the 1970s.
AI is useful for several tasks associated with medical image analysis.
Techniques for Medical Image Analysis
Radiologists and AI systems must segment images to identify biologically relevant components, such as bones, organs, and tumors. Graph models, such as Markov random fields (MRF) based on Bayesian statistics, have been used to assess multiple sclerosis lesions in human brains.
Clustering techniques are also useful for segmentation problems and have the advantage of lower computational complexity than graph models. Researchers segmented brain images using spectral clustering to analyze the structure of the thalamus in human brains.
Diagnosticians often work with multiple images created by different scanning technologies. Integrating these images, in a process known as image registration, can help improve diagnosis.
There are several approaches to this task, including mutual information-based registration. This technique builds on mutual information across images to align and coordinate images. Mutual information is a building block of other AI systems, such as some text classification systems.
Once images are segmented and registered, diagnosticians can delve into the core tasks of diagnosing and detecting medically significant elements of the images, sometimes with the support of computer aided diagnosis (CADx) and computer aided detection (CADe) systems.
These support tools build on several types of AI methods, including pattern recognition, machine learning, and knowledge representation. For example, researchers created a classifier to identify components of human lungs from CT scans of the chest using the ADABoost classification algorithm.
Deep learning is especially important to recent advances in medical image analysis. Convolutional networks enable highly accurate image classification while maintaining tolerable training time requirements. Deep learning techniques are the foundation of commercially developed medical image analysis and decision support systems, such as AIDoc.
Bringing AI to Clinical Settings
AIDoc takes an integrative approach to diagnosis. The system’s deep learning network is trained to recognize multiple types of anatomical objects but the complete analysis platform includes other clinical data about the patient as well.
The AIDoc system integrates into existing picture archiving and communications systems (PACSs). This is an important consideration, since the core AI diagnostic algorithms are only useful if they are deployed in ways that streamline and enhance the image diagnosis problem.
By bringing deep learning and other AI techniques into existing workflows, AIDoc avoids disrupting existing procedures or creating siloed tools that less likely to be used.
Behold.ai is another company that is leveraging artificial intelligence to improve the quality of image analysis while driving down costs. They are currently focused on working with radiologists in the United Kingdom’s National Health Service (NHS).
AI techniques can help improve the speed and quality of medical image diagnosis but there are limiting factors we must keep in mind. Image analysis classifiers, especially deep learning models, require large amounts of data.
Privacy regulations may limit or slow access to images and metadata associated with those images delaying the data acquisition phase of model development. Non-deep learning techniques may be able to build models with less data, but many of these are not effective for medical image diagnosis.
Like other AI systems, they must be developed and deployed with an eye to humans who will ultimately use and depend on them. AI must be available through existing tools and workflows, and a diagnosis must be presented in ways that human experts can assess.
This can be difficult for deep learning and other neural network models, which are generally less transparent than other machine learning methods, such as decision trees and random forests.
Medical imaging is a valuable diagnostic tool but interpreting images requires specialized skills. AI techniques are are increasing the volume of images human experts can review enabling wider use of imaging and improving the quality of care.