Contributions to the standarization and optimization of hyperspectral microscopic systems for medical diagnosis

Abstract

Hyperspectral (HS) microscopy has emerged as a promising modality in biomedical imaging, enabling the capture of highresolution spatial and spectral information from biological tissues. Its ability to provide non-invasive, label-free, and quantitative analysis of microscopic samples opens new possibilities in digital pathology and computer-aided diagnosis. HS microscopy retains critical biochemical and morphological features in tissue specimens, allows for retrospective reanalysis, and generates reproducible highdimensional data. These advantages position it as a strong complement to traditional histopathological techniques. However, despite its potential, the adoption of HS microscopy in routine clinical workflows remains limited. Current barriers include significant variability in instrumentation, imaging protocols, and data processing techniques, which challenge reproducibility, scalability, and clinical integration. Overcoming these limitations requires a holistic and technically rigorous approach, starting with the development and characterization of the imaging instrumentation itself. Reliable and consistent data acquisition depends on thorough evaluation of the sensor’s performance, including spectral accuracy, spatial resolution, illumination uniformity, and noise characteristics. Establishing these parameters ensures fidelity and comparability across imaging sessions and platforms, setting the stage for robust downstream analyses. Another critical consideration is the effect of optical focus on image quality. HS data is particularly sensitive to even slight defocusing, which can distort spectral signatures and compromise interpretability. Strategies such as focus quality assessment and optimization routines—incorporating sharpness metrics, autofocus evaluation, or focus stacking—play a vital role in maximizing both spatial and spectral integrity across entire tissue sections. Equally important is the standardization of sample preparation procedures. Variability in histological processing can introduce inconsistencies that obscure genuine spectral patterns. In particular, tissue thickness has a notable impact on spectral quality and contrast. The adoption of well-defined sectioning protocols, informed by empirical evidence, is essential for minimizing intersample variability and enhancing reproducibility, especially in studies involving multiple centers or longitudinal datasets. Finally, once data are acquired, specific preprocessing steps must be applied to prepare the hyperspectral information for meaningful analysis. These include background subtraction, spectral normalization, denoising, and illumination correction, operations tailored to the unique properties of HS microscopy data. Properly implemented, these workflows mitigate acquisition artifacts and strengthen the reliability of extracted spectral signatures. Complementing these technical advancements, this Ph.D. thesis presents the development of curated, publicly available HS microscopic databases designed to capture real-world variability in sample preparation and imaging conditions. These benchmark datasets provide a valuable foundation for reproducible algorithm development and foster collaboration between computational researchers and clinical practitioners, addressing a critical bottleneck in the field and accelerating the translation of HS microscopy into clinical practice. Together, these efforts contribute to the maturation of HS microscopy as a reproducible, scalable, and clinically relevant imaging modality. By emphasizing standardization, reproducibility, and data quality throughout the imaging pipeline, this Ph.D. thesis lays the groundwork for future diagnostic assistance tools based on HS microscopic technology. These developments are expected to facilitate the integration of HS microscopy into both clinical research and practice, supporting more objective and precise diagnostic decision-making in biomedicine.