A Promising Contribution of Hyperspectral Unmixing for Overlapped Tissues Segmentation Using Autoencoders

Accurate segmentation of overlapping biological tissues is recognized as a major challenge in medical imaging, particularly in computer-assisted surgeries. Traditional methods are based on binary masks that indicate whether a material is present or not in a pixel. However, during surgical procedures, tissues are often overlapped, for example, when blood covers other structures resulting in mixed rather than pure pixels. In such cases, binary masks are rendered ineffective and the surgeon's view is complicated. This difficulty is referred to as the mixed pixel problem. In this study, a novel approach is introduced, representing the first application of hyperspectral unmixing in the medical domain to address this limitation. An unsupervised deep learning framework, using a pixel-based autoencoder, is employed to unmix hyperspectral images of key biological tissues, including bone, fat, blood, and cartilage, in both pure and overlapping regions. For each pixel, the proportion of each tissue is estimated based on its unique spectral signature, and continuous abundance maps are generated to replace conventional binary masks in order to accurately represent both pure and mixed pixels. Through this approach, precise identification of tissue types is achieved by extracting their spectral signatures and corresponding abundances, thereby enabling improved segmentation and visualization of complex tissue structures.