Making the Machine More Intelligent

The automatic scan planning feature on Canon Medical's new Aquilion Serve CT scanner is pushing the boundaries in image acquisition, to let clinicians focus on what really matters: patients. VISIONS spoke with Marco Razeto, Principal Scientist at Canon Medical Research Europe (CMRE) in Edinburgh, UK, about Anatomical Landmark Detection (ALD), the algorithm behind the technology.

Anatomical Landmark Detection is an AI-driven algorithm and one of the most important technologies used in automatic scan planning, according to Marco, who supervised the team responsible for the creation of the tool.

‘ALD identifies specific anatomical locations using a 3D landmark scan - i.e. an ultra-low dose CT scan of the whole body, and names them,’ he explained.

ALD marks the location with a particular name code so that the machine knows where parts of the body are in space. ‘This way, you can set boundaries for your scan and don’t have to touch anything else. You can focus on exactly what you want to image.’

The tool improves workflow, increases clinicians and technicians confidence, and improves consistency between radiographers regardless of their experience. Radiographers no longer have to take 2D scanos and drag boxes around the area they want to image - a process that can take up to 20 seconds for an experienced technician.

‘You save time and energy,’ he said. ‘The clinician and the technician can care more about the patient and less about planning the examinations.’

With automatic positioning, the risk of making a mistake is also drastically reduced, as the system doesn’t get tired or distracted.

The main benefit of ALD is that it enables the machine to understand what it sees and makes it more intelligent, Marco believes. ‘The machine can see and name the body parts it’s looking at. You can tag an image to say what’s in it, and you can pre-process the image to be used for a clinical application. The machine can name the content of the image. It gives you automatically fulfilled information.’

Initial works for the creation of ALD started over a decade ago, to answer a concrete clinical question. ‘Scientists wondered if we could find an anatomical location directly from 3D images,’ he recalled. ‘We developed the tool over many years for different purposes, including aligning images of the same patient over time on different modalities.’

About four years ago, the CMRE team was invited to adapt the algorithm to automatically determine the borders of a CT scan.

Over 30 scientists have worked to refine the algorithm for this particular setting. ‘It’s really a collaborative effort between everyone at CMRE, Canon Medical Europe and Canon Medical Japan,’ he said.

The team initially trained the algorithm on hundreds of manually annotated CT images. For the automatic scan planning feature, they did additional training on hundreds of 3D landmark scans, since the dose and appearance are different from routine axial CT images.

The researchers at CMRE were among the first to develop a solution using AI and faced a number of issues at the time. ‘We solved most problems before the explosion of machine learning and deep learning,’ said Marco, a graduate physicist and computer engineer, who helped pioneer image analysis powered by AI and other computational methods 20 years ago. ‘We were exploring a new technique, so there wasn’t much understanding on the market yet.’

Another problem for these early developers was the lack of availability of the right data. ‘That’s an issue with every machine learning algorithm. Even today, it’s hard to explain that we need a mountain of data to be able to do these things.’

Furthermore, the product is based on a particular type of imaging which was available only at one or two sites in the world back then, slowing down data collection.

Nevertheless, the team found a way to tackle these issues and managed to test the tool with quantitative analysis of the results. ‘We checked not only whether we identified the landmarks we wanted but also how far they are from their original positions - the right place. We test the tool continuously.’ ALD’s detection accuracy is in the 2mm range, making it very reliable to work with in daily practice.

Further improvements of ALD will focus on making it faster, better and, possibly, extending it to other modalities, depending on the clinical needs.

A low-hanging fruit with the technology is the ability to do cross-modality imaging, as ALD can be adapted to MRI or ultrasound, Marco believes. ‘That could be very interesting for patients with cancer or chronic diseases that require life-long imaging,’ he concluded.

Marco Razeto is a Principal Scientist at Canon Medical Research Europe (CMRE). Marco graduated in physics in the late 1990s at the University of Genoa, Italy. He then moved to the UK, where he completed his PhD in computer engineering, computer vision and medical imaging in 2005 at Heriot-Watt University, Edinburgh. He started working for Voxar, which was later acquired by Barco and then Toshiba Medical Systems, which Canon Medical acquired in 2016.

Canon Medical Research Europe (CMRE) in Edinburg, UK is a 130-strong team working on improving and automating patient care workflow, right from the start of the patient journey - from image acquisition to diagnosis, treatment and monitoring. CMRE is a centre of excellence for the development of Machine Learning technology and half of the team is working on developing new technologies.