Eye and the scalpel

2 April 2020:

Australia is the sunniest continent on Earth — which is why it also has the highest rates of skin cancer. But plentiful sunlight is also likely responsible for the lesser known ‘ocular surface cancer’, which occurs when abnormal cells on the eye grow and divide in an uncontrolled way.

But eye tumours can also be benign — hence, knowing if a growth is malignant or not is crucial. To be certain, an ophthalmologist needs to cut into the eye with a scalpel to take a sample, or biopsy, of the affected area for testing.

‘You can imagine how unpleasant it would be if someone has to have a piece of your eye cut out for a biopsy,’ said Dr Abbas Habibalahi, a postdoctoral researcher at the Centre for Nanoscale BioPhotonics (CNBP) node at the University of New South Wales in Sydney. ‘Afterwards, you need an eyepatch or have blurred vision for a few days, and have to use eyedrops for weeks. Then there’s redness, swelling and the risk of infection.

‘So, it’s invasive — but biopsies have other problems,’ he added. ‘One is that it can take about one week for pathology to confirm. Another is that the surgeon has to sample the tumour just right, as they may take the sample from an area that is not cancerous and get a false result.’

That’s likely to change. CNBP has developed an automated, fast and effective way to diagnose a type of eye surface cancer, known as ocular surface squamous neoplasia or OSSN. It relies on shining LED lights with a number of distinct wavelengths at eye tissue: cells in the eye then absorb the light energy and emit a natural glow, known as ‘autofluorescence’. This can be used to distinguish cell structures from each other, since the chemical composition of each structure varies, and this variability shows up in the reflected glow.

A small and highly sensitive camera captures the autofluorescence, and the images are cycled through a computer that uses artificial intelligence to recognise diseased and non-diseased eye tissue, in real time. Because one of the defining characteristics of cancer cells is that they multiply much more rapidly than surrounding tissue, their metabolites — small molecules that power things like growth and signalling — are very different, and give them away.

‘We’ve developed not just the technology to scan the eye, but the methodology to identify the biochemistry inside the cells — metabolites like NADH [nicotinamide adenine dinucleotide and hydrogen], porphyrin and flavins — and the difference in their concentrations,’ added Dr Habibalahi, whose work is led by Prof Ewa Goldys, CNBP’s deputy director.

Having previously established how the autofluorescence from ocular surface cancer cells is different from normal eye cells the CNBP researchers built an artificial intelligence classifier that can now instantaneously recognise cancer tissue. A multidisciplinary approach was essential, requiring engineers, ophthalmologists, pathologists and biologists to bring it to fruition.

What’s really exciting is that the technology is excellent at determining the boundary of the ocular surface cancer. This allows an ophthalmologist — when removing a tumour — to know exactly where the diseased tissue is. Currently, removing a tumour usually requires the removal of surrounding healthy eye tissue, in case unseen tumour cells are left behind. Despite this, reoccurrence rates of ocular surface cancer can be as high as 40%.

‘There is no way to identify the boundary of the cancer, yet knowing where the cancer ends is very important,’ said Dr Habibalahi. ‘Eye tissue is very precious, so the surgeon doesn’t want to sacrifice healthy cells.’

In 2019, a trial on 18 patients with tumours from OSSN successfully distinguished between diseased and healthy cells, producing a map of the ailment, and matched pathology tests. Next steps are to train the artificial intelligence classifier to identify the various types of ocular surface cancer present by looking through many samples of confirmed disease, and begin on the long road to getting the technology approved for use in clinics.

Related paper:

Title: Optimized Autofluorescence Spectral Signature for Non-Invasive Diagnostics of Ocular Surface Squamous Neoplasia (OSSN)

Journal: IEEE Access

Authors: A. Habibalahi, A. Allende, C. Bala, A. G. Anwer, S. Mukhopadhyay and E. M. Goldys

Abstract: Clinical OSSN diagnostics by non-invasive spectral imaging of eye autofluorescence must be rapid enough to be comfortable for patients – without compromising accuracy. This requires identifying optimized spectral signatures of OSSN based on a minimal number of spectrally defined images. Here, we identified such signatures using a data-driven methodology of swarm intelligence. Ten patients with histopathological diagnosis of ocular surface squamous neoplasia (OSSN) were recruited. Their unstained biopsy OSSN specimens were investigated using a custom-built autofluorescence multispectral microscopy imaging system. The images were taken in 38 spectral channels spanning specific excitation (340 nm-510 nm) and emission (420 nm-650 nm) wavelength ranges. To identify optimized spectral signatures of OSSN from a small number of channels, swarm intelligence was combined with discriminative cluster analysis. This study established an optimized spectral signature of OSSN derived from multispectral data taken in 38 channels. Depending on the critical nature of the application and the consequences of misclassification error, two optimized spectral signatures with 5 and 10 channels were obtained which reduced the imaging time to 20 and 40 seconds, a reduction by 75% and 80 %, respectively. The K-nearest neighbor classifier was then built using OSSN spectral signatures and optimized to successfully detect OSSN with ~1% and ~14% misclassification error using 10 and 5 channels, respectively. Our study found an optimized spectral signature of OSSN allowing rapid diagnostic imaging in clinical settings and demonstrates the feasibility of using optimized multispectral autofluorescence spectral signatures to detect and determine boundaries of OSSN.

Linkhttps://ieeexplore.ieee.org/abstract/document/8846190/authors#authors