Multispectral Imaging, Image Analysis, and Pathology

Abstract

Biological systems are complex; multiparameter detection methods such as expression arrays and flow cytometry make this apparent. However, it is increasingly important not just to measure overall expression of specific molecules, but also their spatial distribution–at various scales and while preserving cellular and tissue architectural features. Such high-resolution molecular imaging is technically challenging, especially when signals of interest are co-localized. Moreover, in fluorescence-based methods, sensitivity and quantitative reliability can be compromised by spectral cross-talk between specific labels and also by the presence of autofluorescence commonly present, for example, in formalin-fixed tissues. In brightfield microscopy, problems of overlapping chromogenic signals pose similar imaging difficulties.

These challenges can be addressed using commercially available multispectral imaging technologies attached to standard microscope platforms, or alternatively, integrated into whole-slide scanning instruments. However, image analysis is a central and still incompletely solved piece of the entire imaging process. New and evolving machine-learning technologies as well as other image-understanding approaches can create tools that can readily be used to separate image regions into appropriate classes (“cancer”, “stroma”, “inflammation”, e.g.) with (near) clinically acceptable accuracy. By itself this is useful, but can also be combined with specific segmentation and quantitation tools to extract molecular data automatically from appropriate cellular and tissue compartments, information necessary for designing and testing targeted diagnostic and therapeutic reagents.

Having tools such as these available will allow pathologists to deliver appropriate quantitative and multiplexed analyses in a reproducible and timely manner.