AI-powered spinal segmentation technology could ‘save radiologists and surgeons thousands of hours of work’

By March 31st, 2021

Spinal segmentation CSIRO's Data61

An example of spinal segmentation using this new technology.

A new AI-powered piece of medtech that segments spinal vertebrae within two minutes with an accuracy rate of 95% could enable medical professionals to better plan their surgeries and design custom implants. 

Designed by CSIRO’s Data61 in collaboration with Australian Medical Imaging Company Singular Health, this automated segmentation and labelling technology is intended to be used in tandem with Singular Health’s MedVR & 3DicomViewer software to visualise and manipulate the spine and individual vertebrae in 3D.

Previous segmentation methods required thousands of hours of manual identification and mark ups to computerised tomography (CT) scans, of which there can be hundreds.  

Manual segmentation is a hugely labour-intensive process, however, by automating it using AI, users will not only save time and energy, but ensure a high degree of segmentation and localisation accuracy,” explains Data61 research lead, Dr Dadong Wang.  

Radiologists and surgeons will now only have to make small mark-ups and validate the software’s results, says Dr Guan Tay, Executive Director of Innovation at Singular Health. 

“This semi-automated segmentation process will allow surgeons and radiologists to edit the output and ensure 100% compliance with their interpretation of the image whilst still saving vast amounts of processing time,” he explains.  

“The use of artificial intelligence in medical imaging, and more specifically radiology, has the ability to profoundly change the workflow for radiologists.” 

Spinal segmentation CSIRO's Data61

An example of spinal segmentation software.

Using a dataset of more than 200 CT scans of labelled data, the Data61 team trialled a number of different AI models and pre-and-post processing techniques to achieve the instance segmentation, labelling, surface meshing and spatial location of the individual vertebrae. 

The AI was developed by adapting deep learning based instance segmentation approaches such as nnUNET, SC-NET, and Dense-NET,” says Dr Wang. 

The models were trained using a public dataset called VerSe’2020, which included 100 CT scans of spines collected from people of different age and gender.  

The trained models were then tested on another 100 CT scans to generate segmented labels of spine, individual vertebrae, spatial location and identification of each vertebra.”  

The technology will be integrated into Singular Health’s MedVR software, which is currently offered to hospitals, individual clinicians, schools and universities 

The project was made possible through CSIRO Kick–Start, an initiative that provides funding and support for innovative Australian start-ups and small businesses to access CSIRO’s research and development (R&D) expertise and capabilities. 

0 comments

Leave a Reply