logo
Project Brochure |  
Project Outcomes
 

HELICoiD Demonstrator In-Theatre

The hyperspectral demonstrator system developed by HELICoiD was used during three extended in-theatre trials. It was first deployed to the Hospital Dr. Negrín at Las Palmas de Gran Canaria from March to October of 2015, and a second trial was conducted at the University Hospital of Southampton, UK, between October of 2015 and May of 2016. The demonstrator was then returned to the University Hospital Dr. Negrín for final experiments between June and December of 2016. In total, 93 hyperspectral datacubes, showing a range of primary and secondary tumours, were generated during 36 surgical procedures.
 
 
 
Demonstrator at the University Hospital Doctor Negrín of Las Palmas de Gran Canaria Image
 

Acquisition Protocol

Images were captured after duratomy before the arachnoid and pia had been breached. If the tumour could be seen on the surface, two sterilised rubber ring markers were placed to identify the position of the tumour as well as a region of normal brain tissue (based on visual appearance, anatomical relationship to sulci and gyri and image guidance feedback). Biopsies taken from inside the rings then allowed the images of these marked regions to be compared with results from pathology.
 
 
 
RGB representation of hyperspectral image with the markers placed over the brain tissue surface and tumour marked in yellow.
 

Algorithms for Tissue Classification

The final algorithm developed by HELICoiD combines both supervised and unsupervised learning techniques. Unsupervised learning approaches allow clustering of pixels automatically by similarity and were found to reliably create groups of pixels corresponding to a single kind of tissue or material. However, it is not then possible to directly perform a medical diagnosis, this step requires use of supervised techniques. While supervised learning algorithms are an effective approach to tissue classification due the ‘capture’ of the neurosurgeon’s expertise, the accuracy of their predictions nevertheless depend strongly on the quality of the labelled data used as the ground truth. This ground truth data is expensive to obtain, requiring time-consuming involvement of experts in the manual classification of groups of image pixels by tissue type. An interactive graphical user tool was therefore developed to streamline this training and labelling process.
 
The classification output of the algorithm can be shown as a map, where each colour corresponds to different types of tissues and other materials present. Using hardware acceleration developed by HELICoiD, the time to process a cube decreased from approximately 10 minutes to 1 minute (depending on the image size). The final algorithm was applied to 15 hyperspectral data cubes, demonstrating strong correspondence with the known tumour locations.
 
 
 
Density map after classification using HELICoiD algorithms.