Description
Scientific context and objectives Histopathology remains, to this day, the gold standard for diagnosing skin cancer, based on a surgical biopsy taken from the patient followed by laboratory examination of the sample. To improve the diagnostic efficiency of these biopsy procedures, optical spectro-imaging methods – which rely on the optical characterization of healthy and pathological tissues – have been under development for many years, given their non-invasive nature, their sensitivity to morpho-functional variations in biological tissues, and their applicability in vivo. To support the clinical roll-out of these approaches (intraoperative diagnostic assistance, surgical guidance), the development of reliable and robust methods for analyzing datasets derived from the various optical biopsy modalities is a major challenge. The proposed postdoctoral research focuses on the development and validation of algorithms for the processing and analysis of in vivo spectroscopic data, whilst also exploiting information obtained from ex vivo tissues via complementary spectro-imaging modalities. It relies, on the one hand, on a set of spatially resolved diffuse reflectance (DR) and multi-excitation autofluorescence (AF) spectra acquired in vivo in clinics using the SpectroLive device [5,4], and, on the other hand, on various datasets that have been or will be acquired as part of the Opticskin ANR project on the same skin tissues ex vivo, namely hyperspectral microscopy (HSM) and fluorescence microscopy (FM) images, Raman microscopy (RM) images, Mueller polarimetric microscopy (PM) images, and fluorescence lifetime-resolved microscopy (FLIM) images. The following complementary research axes will be considered: 1. Acquisition and segmentation of HSM & FM images: The histological classification resulting from the pathological examination of samples does not reflect the complex structure of the lesions and their environment within the tissue volume probed by in vivo optical spectroscopy, leading to uncertainties regarding the reference classification. Conversely, HSM, FM, MRI, PM and FLIM images of tissue sections enable improved spatial mapping, paving the way for more reliable labelling based on quantifiable criteria, as well as more precise characterizations (stages of disease, risk of recurrence, etc.). This part of the work will focus on the implementation of the hyperspectral microscope instrumentation systems at CRAN’s PhotoVivo platform for the acquisition and segmentation of HSM and FM images of the available skin sections. This work will also involve mastering metrological procedures and defining experimental protocols using optical phantoms and biological tissues. The images will then be segmented in order to highlight certain properties of pathologies at the microscopic scale. For this task, we will draw inspiration from deep neural network-based approaches proposed for PM images, for example [2]. 2. In vivo classification using optimal transport: The skin lesions in the SpectroLive database are divided into three main classes: squamous cell carcinoma (SCC), basal cell carcinoma (BCC) and actinic keratosis (KA). Each lesion is characterized by a set of AF and DR spectra, measured on the lesion itself, on perilesional areas, and on healthy reference areas. Supervised classification approaches tested on the spectra acquired with this device [1,6,7] still struggle to distinguish between the different classes of lesions, suffering in particular from the high heterogeneity of the data, both between and within individuals. A recent study [3] has shown that Wasserstein distances [8] may offer an advantage over standard metrics for unsupervised spectrum clustering, but remain nonetheless sensitive to this heterogeneity. In this thesis, we will consider classification methods at the lesion level rather than at the level of individual spectra, based on the Gromov-Wasserstein distance, which takes into account the relationships between spectra within the same group (i.e. the same lesion). In particular, we will focus on the ‘fused Gromov-Wasserstein’ extension [9] in order to establish a characterization of lesions based both on the individual properties of the spectra and on the structure of the groups to which they belong. 3. Fusion and multimodal classification: In order to exploit multimodal information to guide in vivo classification, segmentation information extracted from the various modalities—starting with RM images (already acquired), HSM and FM—will be fused into a single multimodal label, based on state-of-the-art deep learning approaches in the field of data fusion [10]. These multimodal maps will then be used as new labels and integrated into the optimal transport classification approaches described above. The final in vivo classification will thus also take into account ex vivo information to distinguish between different lesions and infer their stage of progression. Performance in distinguishing between the three classes of pathology will be compared with histological classification and also analyzed in terms of the contribution of each optical method. The scientific output will include the generation of data, publications in national and/or international journals, and participation in national and/or international conferences. References [1] F. Abdat et al. Opt. Express 20.1 (2012), pp. 228–44 [2] S. Chae et al. arXiv:2602.09787, 2026. [3] A. Dieng et al. Proc. SPIE Photonics Europe 14098(2026) [4] T. Elsen et al. Data in Brief 53(2024), p. 110163. doi: https://doi.org/10.1016/j.dib.2024.110163. [5] T. Elsen et al. SpectroLive : optical spectra (autofluorescence and diffuse reflectance) acquired on human skin carcinomas. 2023. doi: 10.12763/EYVX3P. url: https://doi.org/10.12763/EYVX3P. [6] V. Kupryanov et al. J. Biophotonics 16.7 (2023) [7] D. Ni et al. ECBO Technical Digest, M3A.4(2025) [8] G. Peyré and M. Cuturi.Foundations and Trends in Machine Learning 11.5-6 (2019), pp. 355–607. [9] T. Vayer et al.Algorithms 13.9 (2020) [10] Y. Zhang et al. Image and Vision Computing 105 (2021), p. 104042.
