Hyperspectral Unmixing for Orthopedic Tissue Identification Under Spectral Variability and Material Overlap.

Enhancing surgical precision through automated tissue identification requires overcoming the complexities of real operative scenes. Specifically, material detection is complicated by tissue superposition (e.g. thin layer of blood masks underlying structures), intricate anatomical geometries, and lighting instability. These factors manifest as the mixed pixel and spectral variability problems, which remain the foremost challenges for hyperspectral imaging (HSI) in clinical applications.In this context, this paper presents a feasibility study leveraging Visible and Near-Infrared HSI combined with hyperspectral unmixing algorithms to address these complexities. Ex-vivo bovine samples, comprising key orthopedic tissues such as blood, ligaments, cartilage, bone, and fat, were imaged using HSI in a manner that replicates real operative challenges. Subsequently, six state-of-the-art unmixing algorithms selected for their ability to simultaneously resolve mixed pixels and mitigate spectral variability effects are evaluated. Through a comparative analysis, model-based unmixing algorithms demonstrated optimal performance in producing abundance maps that accurately quantify the proportion of distinct tissue types, even within a single pixel subject to variability perturbations. These algorithms successfully decouple environmental variability, such as shadowing, illumination shifts, and shape spectral deformations, from actual material proportions, preventing altered spectral signatures from confusing true abundance quantification.