A new segmentation framework for infrared spectroscopic imaging using frequent pattern mining

Histologic analysis of a stained tissue sample by a trained pathologist forms the definitive diagnosis of prostate cancer. Rapid and objective second opinions are highly desirable to make more accurate diagnostic decisions. One alternate method is to use Fourier transform infrared (FT-IR) spectroscopic imaging, which is an emerging technique that combines the molecular selectivity of spectroscopy with the spatial specificity of optical microscopy. While instrumentation is well-developed for FT-IR imaging, information extraction from the data could benefit greatly from improved approaches. Here we propose a new approach to segment histologic classes in a tissue for FT-IR imaging using frequent pattern mining. Prior to applying frequent pattern mining, FT-IR images are discretized, and subsequent pruning method and feature selection method result in a classifier for the segmentation. The method is evaluated using two different datasets. Results indicate that accurate histologic segmentation is achievable by this approach.