Skin cancer margin detection using nanosensitive optical coherence tomography and a comparative study with confocal microscopy

Excision biopsy and histology represent the gold standard for morphological investigation of the skin, in particular for cancer diagnostics. Nevertheless, a biopsy may alter the original morphology, usually requires several weeks for results, is non-repeatable on the same site and always requires an iatrogenic trauma. Hence, diagnosis and clinical management of diseases may be substantially improved by new non-invasive imaging techniques. Optical Coherence Tomography (OCT) is a non-invasive depth-resolved optical imaging modality based on low coherence interferometry that enables high-resolution, cross-sectional imaging in biological tissues and it can be used to obtain both structural and functional information. Beyond the resolution limit, it is not possible to detect structural and functional information using conventional OCT. In this paper, we present a recently developed technique, nanosensitive OCT (nsOCT), improved using broadband supercontinuum laser, and demonstrate nanoscale sensitivity to structural changes within ex vivo human skin tissue. The extended spectral bandwidth permitted access to a wider distribution of spatial frequencies and improved the dynamic range of the nsOCT. Firstly, we demonstrate numerical and experimental detection of a few nanometers structural difference using the nsOCT method from single B-scan images of phantoms with sub-micron periodic structures, acting like Bragg gratings, along the depth. Secondly, our study shows that nsOCT can distinguish nanoscale structural changes at the skin cancer margin from the healthy region in en face images at clinically relevant depths. Finally, we compare the nsOCT en face image with a high-resolution confocal microscopy image to confirm the structural differences between the healthy and lesional/cancerous regions, allowing the detection of the skin cancer margin.


Other tissue sample nsOCT results
We show another two samples of nsOCT results for further structural comparison of the healthy and lesional regions with an intervening marginal area. These two samples were collected from the nose and scalp of human skin. , it can be clearly observed that the lesional region to healthy region mean spatial period (H z ) value changed from 482 ± 1.28 (mean SP ± standard deviation) nm to 475.68 ± 1.37 nm. Corresponding Fig (f) show the changing of the mean spatial period from 473.68 ± 1.01 nm to 479.71 ± 0.98 nm. Red dashed lines in figs. (c) and (f) represent the mean spatial period values of each region. As a result, the mean spatial period of the structures has increased from the healthy region to the lesional region. All three samples followed the similar mean spatial period profiles trained.

Tissue sample overview from OCT en face image, confocal image, and H&E slide image
To get an overview of the tissue sample, parallelly we have shown three different system images with the same sample orientation. Fig. (a) represents a conventional intensity-based OCT en face image. As shown in Fig (a), intensity-based OCT en face image cannot discriminate between healthy and lesion regions. Fig. (b) represents a laser scanning confocal microscopy image of the same tissue block. The right side of the image appears to lesion region and the left side is the healthy region. 20X objective lens with an 0.75 NA (Numerical Aperture) has been used for confocal microscopy imaging. Fig. (c) represents a low-resolution H&E slide image to get an overview of all the regions of the same tissue. To acquire this H&E slide image, a 4X objective lens with 0.13 NA has been used to image healthy and lesion regions.

Histology image analysis results
We used an Olympus light microscopy of 580 nm excitation wavelength and an objective lens with an NA (Numerical Aperture) 0.75 for acquiring the H&E slide image. Image consisted of 2056 X 1920 pixels covering an area of 283 X 212 mm, which corresponds to 110 nm/pixel with a lateral resolution of 286 nm. The light microscopy image of the H&E slide is presented in Fig. 4(a). To validate the nsOCT results with the light microscopy image, spatial frequency changes of the structures within the two regions have been calculated. To analyze the spatial frequency profiles black square boxes regions were taken of the light microscope image 4(a). We have analyzed two hundred vertical line profiles from the black square boxes. To obtain the spatial period distribution of the structures, Fourier transforms have been applied on these line profiles. Median spatial frequency/period (MSP) of the spectrum was calculated after the Fourier analysis of the spectrum profiles. Figure 4(b) box plot shows the distribution of the mean spatial periods of healthy and lesion regions, and it is calculated from the line profiles of fig 4(a). The box plot indicates that the spatial period of the nanoscale structures within the lesional region increased compared to the healthy region. The mean spatial period distribution of the healthy region is 837 ± 123 nm (mean SP ± standard deviation) and for the lesional region is 1000 ± 151 nm.

Confocal and histology image analysis flowchart
In order to analyse the spatial frequency profiles from the confocal and H&E slide images, we analysed two hundreds vertical line profiles across the confocal image in a given region and Fourier transform of these profiles were calculated to obtain the spectrum of spatial period distribution of the structures as shown in flowchart below. The regions within the confocal image were carefully chosen to avoid any artefacts caused by fixation/folding. From the Fourier spectrum of the profiles, median spatial frequencies/periods of the spectrum were calculated using Eq. S1.
where x(n) represents the intensity of the spatial frequency/period profile at the frequency/period f(n). This method characterizes the median local structure size within the sample.

nsOCT image formation
In order to realize nsOCT methods, we first obtain the processed interference spectra after basic preliminary modifications, including k-space linearization, background noise removal, apodization, and dispersion compensation. Then the k-space linearized spectral interferogram I(λ) is converted to corresponding axial spatial frequency I(ν z ) using Eq. (2). Next the processed spectra of axial spatial frequency are decomposed into several sub-bands (N=10 number of subbands) using Tukey windows as for nsOCT imaging. We can use different window widths based on the sample and imaging aim to improve structural or spatial resolution. For each of the N zones, the axial spatial frequency profile is inverse Fourier transformed to reconstruct the OCT image for each zone. Zero-padding was used before FFT for each zone to better localize the peak frequency. From the reconstructed OCT images of N zones for each point of depth profiles we form an energy contribution versus the corresponding spatial period of the sub-band plot, called the spatial frequency/period profile. Then, the nsOCT image can be formed as a color map of some informative parameters of these spatial frequency profiles which better visualize the structural changes. In this paper the dominant spatial frequency / period value which corresponds to maximum energy contribution is mapped to create the nsOCT image. This procedure is repeated for each A-line in a B-scan, as well as for each B-scan in a 3D volume. After the Fourier transformed from each zone, we subtracted the noise using thresholding. Furthermore, a 4 x 4 spatial kernel and a threshold value of 1.8 was applied to suppress noise within the nsOCT images for this application. Depending on the application or purpose, the spatial filtering kernel and the threshold can be optimized.