Challenges and opportunities of building machine learning solutions on the ultrasound machines

Ultrasonography is the most ubiquitous medical diagnostic device, which has been around for more than half a century now. There have been incremental improvements in the hardware and software with devices getting smaller, faster, and more capable. The advent of deep learning and the possibility of using a diagnostic algorithm for sonographic images opens up new possibilities. There can be four kinds of locales for such diagnostic algorithms, on device, off-device, on-premise, and on-cloud. The most efficient location for performance and ease of access is on the device. I will briefly describe the opportunities and challenges in such deployments.
The first-ever on-device radiology AI solution is developed by GE Healthcare. The portable X-ray unit called Critical Care Suite can read chest X-rays and provide real-time alerts for presence of pneumothorax (a critical condition in which air fills the covering of lungs). The advantages of such systems are self-evident.

Similar possibilities exist for sonography machines, where algorithms operating on the equipment can improve the speed and accuracy of the operator. Solutions are being built to analyze the position and habitus of the patient on a real-time basis, and assist the radiologist or sonologist with relevant surface landmarks of internal organs. There are algorithms that process the images as they are being acquired, annotate, and auto-measure anatomical structures like kidneys. Classification algorithms that can identify and characterize lesions in organs can offer instant diagnostic assistance to the radiologist, reducing the scan time while improving the accuracy of diagnosis. There is also an interesting possibility of pattern-matching of the images acquired during scanning and contextual search of similar images from the institutional image database.
When combined with mobility of portable scanners, the machine-learning algorithms can be a powerful diagnostic tool in low-resource settings and areas with low penetration of healthcare services. One such powerful combination is demonstrated by the Butterfly iQ. It is a convex, linear, and sector ultrasound transducer that connects to an iPhone or iPad with the Butterfly app to serve as a full ultrasound machine. The transducer is a multipurpose probe that uses a silicon chip instead of a piezoelectric crystal, which allows it to function as a multipurpose transducer. Such handheld devices have the potential to transform the way in which primary healthcare services are provided in remote locations. But this device as of now processes images on cloud.

But these potential breakthroughs are fraught with equally tough technical and regulatory challenges. The biggest concern that clinicians and radiologists face while adapting machine-learning algorithms in their practice is the lack of clear regulatory framework for approval and guidance. In India, we still look up to American regulatory body, USFDA, before adopting new technologies. But these regulations in the Indian context are purely self-adopted and not mandated by any authority. It also goes without saying that transposing American guidelines to Indian conditions is also not an ideal situation given the vastly different disease, demographic, and socio-economic profile. In addition, in India all sonography equipment are regulated by PCPNDT Act. The possibility of using machine-learning algorithm by nefarious elements for banned indications will be weighed by the authorities before approving any such applications. But this concern is again favoring the option of installing machine-learning algorithms on device rather than processing them on cloud, as it enables more control over the process if the data stays locally.

The major technical limitation of on-device ML solution is the need for putting high compute power on the device than what might be needed for the routine functioning on the equipment. The service and maintenance cycle for the hardware component needed for running deep learning algorithms as of today is vastly divergent from that of the hardware component of the ultrasound equipment. But given the pace at which the hardware and software advancements are progressing, these technical challenges are bound to be solved in a few years. The regulatory challenges need to be addressed swiftly by consensus among all stakeholders to leverage the benefits of the new-age machine-learning revolution.