Multimodal Optical Spectroscopy and Machine Learning For Human Skin Cancer In Vivo Diagnosis

Skin cancers are the most common groups of cancers diagnosed worldwide, among which 90% are non-melanoma skin cancers i.e. Basal Cell Carcinoma (BCC) and Squamous Cell Carcinoma (SCC). Surgery is one of the main steps in treating skin cancer thus resulting in scarring or disfigurement with great impact on patients' quality of life. Skin carcinogenesis modulation factors such as epidermis hyperplasia, collagen enzymatic degradation, overactive cell metabolism, cell pleomorphism, neovascularization etc., are known to modify the optical properties of skin tissues. Thus, the development of non-invasive in vivo optical biopsy methods, to help surgeons in the peri-operative delineation of safety margins, is a major health, social and economic issue.

In the present study, a bimodal system of skin tissue fibered spectroscopy, combining Diffuse Reflectance (DRS) and AutoFluorescence (AFS) measurements, was implemented in the frame of a clinical protocol involving 140 patients with skin carcinomas and Actinic Keratosis (AK). The Spectrolive medical device developed [1,2] includes a multiple optic fiber probe with four source-detector separations from 400 to 1000 µm, a broadband light source for DRS in the spectral range [340-785] nm and five bandpass filtered LED sources for sequential acquisitions of AFS under narrow band excitations between 365 and 415 nm. The clinical protocol consisted in (i) the collection of the clinical examination data (Fitzpatrick chart, Merz scaling, margin delineation), (ii) the spectroscopic acquisition methodology (number and position of the measurements), (iii) the skin tissue resection (surgery) and (iv) the histological analysis of all the excised samples used as reference classification. The multidimensional spectroscopic data set collected [3] was analyzed using a Machine Learning-based approach for automatic supervised classification [4]. Several combination strategies of feature extraction methods (principal component analysis, non-negative matrix factorization, autoencoder), classifiers (support vector machine, linear discriminant analysis, multilayer perceptron, random forest) and data or decision fusion methods (stacking, majority voting) were evaluated. Highest values of accuracy between 83% and 87% were obtained for pair-wise classification comparing BCC and/or SCC vs healthy tissues, with optimized hyperparameters of our classification pipeline.