BCC and Immunocryosurgery scar differentiation through computational resolution‐enhanced OCT images and skin optical attenuation: A proof‐of‐concept study

Monitoring medical therapy remains a challenging task across all non‐surgical skin cancer treatment modalities. In addition, confirmation of residual tumours after treatment is essential for the early detection of potential relapses. Optical coherence tomography (OCT), a non‐invasive method for real‐time cross‐sectional imaging of living tissue, is a promising imaging approach for assessing relatively flat, near‐surface skin lesions, such as those that occur in most basal cell carcinomas (BCCs), at the time of diagnosis. However, the skin's inherent property of strong light scattering impedes the implementation of OCT in these cases due to the poor image quality. Furthermore, translating OCT's optical parameters into practical use in routine clinical settings is complicated due to substantial observer subjectivity. In this retrospective pilot study, we developed a workflow based on the upscale of the OCT images resolution using a deep generative adversarial network and the estimation of the skin optical attenuation coefficient. At the site of immunocryosurgery‐treated BCC, the proposed methodology can extract optical parameters and discriminate objectively between tumour foci and scar tissue.

substantial observer subjectivity.In addition, the relatively strong light scattering properties of the skin are an inherent obstacle to the dermatological application of OCT, as they not only limit light penetration, but also add noise and artefacts to the OCT images.
At the physical level, the strength of the light scattering of a medium is mainly characterized by the optical attenuation coefficient (μOCT), which measures the rate of signal decay verses depth, as light traverses twice in the attenuating tissue. 5Although μOCT is not a substance-specific index, it has found utility in a medical context as an indicator of tissue composition changes that may signal pathological conditions.Immunocryosurgery, a treatment approach that combines topical imiquimod application and cryosurgery, has demonstrated therapeutic effectiveness on par with standard surgical excision for the majority of BCCs.7][8] A distinct finding frequently observed with this therapeutic approach is the timing of recurrences, which appear to occur within the first year of treatment and probably correspond to occult partial responders. 9

| Questions addressed
We conducted a retrospective pilot study to investigate whether an algorithmic approach that extracts skin optical parameters and provides quantitative outcomes, could objectively differentiate between tumour foci and scar tissue in areas where BCCs were treated with immunocryosurgery.

| E XPERIMENTAL DE S IG N
We retrospectively reviewed patient records to identify archival OCT scans of individuals aged 75 years and older, specifically those diagnosed with nodular-type facial BCC and exclusively treated with immunocryosurgery.Inclusion criteria required OCT scans of the BCC lesion before treatment (baseline) and at least two additional scans post-treatment.When available, OCT scans of healthy skin areas, such as the scar periphery or a contralateral site, were also considered as controls.Three tumour areas were included in the study from three patients, two males and one female aged 78-82.All OCT images were acquired in the setting of the Dermatooncology unit of the Univ.Hospital of Ioannina, as part of the routine clinical care for patients with BCC, using the OCT NITID system (Dermalumics, Spain), a portable OCT system with a light source at a central wavelength of 1300 nm, and axial and lateral resolution at 11 Appendix S1).Next, the optical attenuation coefficient, μOCT, was determined on the preprocessed OCT images using the singlescattering model, which assumes that the OCT signal results from light that remains coherent after being scattered only once in the study material (for the theoretical details, see 12 ).Briefly, the model maps the drop in OCT signal power to grayscale values using the exponential decay function of Beer's law.This algorithm is well suited for fixed-focus geometries, which are often used in clinical practice and can also be effectively applied for measurements in samples with layered arrangement of the scatterers, such as in the layers of the human skin. 13Average OCT values were estimated for an array of parallel light decay pathways perpendicular to the epidermis covering an extended z-axis area of the lesion within a coherent tissue ROI of 100 pixels (~ 400 μm) (Figure 1).By applying the canny edge detection algorithm 10 to the SR images, (i) we corrected for the hyperreflective entry point of the epidermis (Ep) in the OCT scans and (ii) the dermo-epidermal junction (DEJ) was identified.The hyperreflective layer formed at the site of light entering the skin does not correspond to the intracutaneous attenuation of light, and if not corrected, it would systematically overestimate the calculated average μOCT values.This effect could be sufficiently mitigated by introducing a small offset of three pixels at the superficial end ('entry') of the light decay pathways to calculate μOCT.In addition, a visually localizable DEJ in the transition between the darker Ep and the lighter dermis signal (D) strongly contradicts the presence of BCC in OCT scans.The Ep-D transition signal manifests itself in the intensity vs. depth diagrams as a characteristic abrupt signal decrease. 14The absence of this finding after immunocryosurgery decidedly indicates ineffective treatment or BCC relapse.

| CON CLUS I ON S AND PER S PEC TIVE S
The data derived from the analysis of the three instructive cases indicate that a focused algorithmic approach to OCT skin imaging holds promise as an objective quantitative computational tool F I G U R E 1 OCT images from three patients: I, II and III.OCT imaging was repeated after 'm' months.Average μOCT was calculated from multiple equal depth light paths (yellow lines).Canny edge detection is marked in red.In some images hyporeflective structures, such as hairs or blood vessels and edge inconsistencies resulted in small ROI gaps.
F I G U R E 2 Relative OCT intensity profiles (in a.u.) plotted against imaging depth (μm).The letter 'm' is used to represent the month in which a subsequent examination was conducted and μ is the fitted attenuation coefficient (μOCT).The second derivative of the signal is displayed beneath each graph.It amplifies signal slope changes, thereby improving the visualization of an Ep-D transition.The shaded regions indicate the Ep-D junction for the cases that it is evident (control and post-treatment skin sites).Below are violin plots of the μOCT values; 3-4 asterisks indicate p-values of p ≤ 0.001 and p ≤ 0.0001; 'ns' denotes non-significant difference.
and 12 μm respectively.The study was approved by the Institutional Review Board of the University Hospital of Ioannina (approval nr: 751/18.09.23) and informed consent has been obtained from the patients involved.The experimental analysis aimed to (i) improve the resolution of the OCT images and (ii) estimate the experimental μOCT in manually defined regions of interest (ROI) on the preprocessed images.Initially, speckle noise was suppressed by using a non-local mean filter employing the scikit-image implementation. 10The non-local means (NLM) algorithm uses intensity and edge information within the ROI to calculate weights, enabling a weighted maximum likelihood estimation of the image output.The algorithm is frequently used in OCT imaging because it tends to smooth out noise while preserving edges and features resulting in a more balanced distribution of pixel intensities.To further reduce the remaining speckle noise of the OCT images, we used the SRGAN technique, 11 a deep learning algorithm from Generative Adversarial Networks (GANs) known for its ability to produce super-resolution (SR) images.The integration of SRGANs represents a new approach that uses 16 blocks deep Residual Neural Network (ResNet) and enables the model to attain a nearly fourfold increase for a wide range of image visuals.With the previous preprocessing steps, we substantially enhanced the quality of the original OCT scans, reducing the grainy pattern of light and dark spots and obtaining a smoother image overall (Figure 1A;

Figure 2
Figure 2 displays the evaluation of the OCT signals of the three included tumours (three panel columns corresponding to the three