This project aims to develop novel algorithms that simultaneously perform spectral unmixing and image denoising to enhance Stimulated Raman Histology (SRH) images. The primary goal is to improve interpretability and reliability for cancer diagnosis while producing high-quality images suitable for downstream deep learning applications, such as computer-aided diagnosis, automatic segmentation, and virtual staining.
SRH is a rapid, preparation-free microscopy technique capable of producing high-resolution histological images of fresh or frozen millimeter-sized biopsies within minutes. It is based on stimulated Raman scattering (SRS) microscopy, a label-free nonlinear optical method. SRH images are acquired at two specific Raman shifts, 2845 cm⁻¹ and 2930 cm⁻¹, corresponding to CH₂ and CH₃ chemical bond vibrations, resulting in two image channels. The CH₂ channel predominantly highlights cell bodies, while the CH₃ channel emphasizes cell nuclei.
Accurate visualization of cell nuclei is critical for cancer diagnosis, as nuclear density, morphology, and spatial organization are key oncological markers. Currently, nuclei are highlighted by simple subtraction of the CH₂ and CH₃ channels. While this provides basic visualization, it often masks subtle structural details and fails to fully exploit the image information. Advanced spectral unmixing approaches, including multivariate curve resolution, independent component analysis, vertex component analysis, and spectral phasor analysis, have been applied to SRS microscopy to assign meaningful spectral signatures to each pixel. However, these methods are generally based on statistical assumptions and do not leverage spatial information.
SRS microscopy is often affected by low signal-to-noise ratio due to laser noise, low target molecule concentrations, or tissue scattering. Both model-based and deep learning denoising approaches have been proposed.
The proposed project aims to develop algorithms that enhance channel independence and increase signal-to-noise ratio while leveraging spatial image content. Two major challenges are addressed: the limited spectral information available from only two Raman shifts and the difficulty of building large annotated datasets due to the emergence of SRH as a clinical imaging technique. The resulting algorithms are expected to produce interpretable, high-quality images that support accurate cancer diagnosis and advanced computational analyses.
SRH is a rapid, preparation-free microscopy technique capable of producing high-resolution histological images of fresh or frozen millimeter-sized biopsies within minutes. It is based on stimulated Raman scattering (SRS) microscopy, a label-free nonlinear optical method. SRH images are acquired at two specific Raman shifts, 2845 cm⁻¹ and 2930 cm⁻¹, corresponding to CH₂ and CH₃ chemical bond vibrations, resulting in two image channels. The CH₂ channel predominantly highlights cell bodies, while the CH₃ channel emphasizes cell nuclei.
Accurate visualization of cell nuclei is critical for cancer diagnosis, as nuclear density, morphology, and spatial organization are key oncological markers. Currently, nuclei are highlighted by simple subtraction of the CH₂ and CH₃ channels. While this provides basic visualization, it often masks subtle structural details and fails to fully exploit the image information. Advanced spectral unmixing approaches, including multivariate curve resolution, independent component analysis, vertex component analysis, and spectral phasor analysis, have been applied to SRS microscopy to assign meaningful spectral signatures to each pixel. However, these methods are generally based on statistical assumptions and do not leverage spatial information.
SRS microscopy is often affected by low signal-to-noise ratio due to laser noise, low target molecule concentrations, or tissue scattering. Both model-based and deep learning denoising approaches have been proposed.
The proposed project aims to develop algorithms that enhance channel independence and increase signal-to-noise ratio while leveraging spatial image content. Two major challenges are addressed: the limited spectral information available from only two Raman shifts and the difficulty of building large annotated datasets due to the emergence of SRH as a clinical imaging technique. The resulting algorithms are expected to produce interpretable, high-quality images that support accurate cancer diagnosis and advanced computational analyses.
Supervisor
Hervé Rigneault, Institut Fresnel, Aix-Marseille Université
Co-Supervisor
Remi Andre, Institut Fresnel, Aix-Marseille Université
Intersectoral partner
Lhoucine Azzi, Lightcore technologies, Marseille, France
International partner
Medical University of Vienna, Austria & Maastricht University, The Netherlands