Hyperspectral materials characterization techniques are becoming ubiquitous and technologically more sophisticated. There has been widespread development and application of machine/statistical learning to better understand large hyperspectral data sets. This has been remarkably fruitful across domains, yet many challenges remain. One important challenge is the utilization of large-scale data sets to uncover generalizable, population-level spatial-molecular differences between materials. We describe the application of mass spectrometry imaging and a novel weakly supervised learning method to identify the spatial-molecular signatures that differentiate two highly similar sets of triple-negative breast cancer, differing only in the expression of bone morphogenetic protein 4 (BMP4). Our method, stability selection-augmented multiple instance learning (SSAMIL): (1) accurately discriminated between the two sets of tumors (92% accuracy); (2) identified the spatial regions differentiating the tumors; and (3) identified the sparse set of molecular signatures used for classification. We validated these signatures using expert pathology annotations and additional imaging modalities, and identified a biomarker that is predictive of metastasis. We also employed high-lateral resolution delayed extraction imaging, identifying the extracellular regions as critical for distinguishing the tumors. This study demonstrates the considerable potential of this approach for medical applications (and, more generally, for materials characterization), in which spatial-spectral characterization is key.
Gardner et al. (Thu,) studied this question.