by ecancer reporter Clare Sansom

The optimum surgical treatment for brain tumours often depends on their histological subtype, but it is generally not possible to determine this until at least some of the tumour has been removed.

The best practice in the care of these patients therefore involves diagnosis during surgery, which is rarely straightforward.

Techniques for intraoperative classification of brain tumours have been under investigation for many decades, but they require complex equipment and skilled pathologists; these are not available in all centres, even in the best-funded health systems.

An ideal system would generate images of tumour tissue for rapid diagnosis and enable variations within the tumour to be assessed and residual tumour in the resection cavity identified and removed during the initial operation.

The optical technique of simulated Raman spectroscopy (SRS) was developed in 2008 and is now used for imaging microscopic tissue specimens without staining.

However, this technique requires two laser pulses to be precisely trained on the specimen and, therefore, complex optics that have proved exceptionally difficult to set up in a working operating theatre.

A large group of scientists and clinicians led by Daniel Orringer and Sandra Camelo-Piragua at University of Michigan Medical School, Ann Arbor, Michigan, USA has now developed and tested a fibre-laser based system of SRS that is small and efficient enough to be implemented within an operating theatre during surgery.

Simply stated, this technique requires two laser beams, one at a fixed wavelength of 790 nm and the other with variable wavelength, to be passed through the specimen simultaneously.

It generates tissue images with a lateral resolution of 360 nm and an axial resolution of 1.8 micometers by mapping two Raman shifts that correspond respectively to the CH2 bonds that are common in lipids and the CH3 bonds that are common in proteins and DNA.

The fibre-coupled laser microscope and all the analytical equipment, which does not require water cooling, was mounted on a single, portable clinical cart for use in the operating theatre.

This microscope generates images of brain tumour and normal brain tissue with similar contrast to those produced by traditional H&E (haematoxylin and eosin) staining and that can therefore be fairly easy to interpret.

The researchers tested their microscope as a diagnostic technique using brain samples from 101 patients that had recently been removed during surgery.

Expert neuropathologists could readily distinguish tumour tissue from normal brain tissue; separate gliomas from non-glial tumours; and diagnose specific tumour types using these images.

The images could also be used to differentiate between sections of tissue with different grades and histology within a single tumour sample.

The researchers used a blinded test involving 30 of the original 101 specimens to prove that the experts could diagnose tumours from SRS images plus medical history and lesion location just as accurately as they could by using H&E stained images instead of SRS ones.

All the pathologists made a few errors using both techniques, and the researchers suggested that these diagnoses might have been correct if larger specimens had been imaged.

This method of diagnosis is not ideal, however, as expert pathologists will not necessarily be available in all hospitals offering surgery for brain tumours.

Orringer, Camelo-Piragua and their co-workers therefore set out to discover whether it would be possible to develop an automatic diagnosis technique using machine learning.

They generated a perceptron – a computer-based classification algorithm – using over 12 thousand field-of-view (FOV) images from the same 101-tumour series, and assigned a set of quantitative attributes to each FOV.

The system was trained to allocate each set of images and attributes to the most likely class of tumour, and iterating until the method gave a recognised correct diagnosis for each tumour.

Finally, the perceptron was tested using a ‘leave-one-out’ approach in which each patient’s images were diagnosed from a ‘training set’ incorporated those from all other patients.

In this test, the algorithm assigned each FOV to one of four diagnostic classes: normal tissue, low-grade glioma, high-grade glioma or non-glial tumour.

The algorithm was able to distinguish tumour from normal tissue with 100% accuracy and to assign the tumour images to the correct diagnostic class in a large majority of cases.

These results suggest that a fibre-laser based stimulated Raman spectroscopy microscope and a neural network-based tool for fast diagnosis can help improve the peri- and post-operative care of brain cancer patients, even when no pathologist is available.

Reference
Orringer, D.A., Pandian, B., Niknafs, Y.S. and 21 others (2017). Rapid intraoperative histology of unprocessed surgical specimens via fibre-laser-based stimulated Raman scattering microscopy. Nature Biomedical Engineering, published online ahead of print 6 February 2017.
doi: 10.1038/s41551-016-0027