Directional Kernel Density Estimation for Classification of Breast Tissue Spectra

Abstract

In Breast Conserving Therapy, surgeons measure the thickness of healthy tissue surrounding an excised tumor (surgical margin) via post-operative histological or visual assessment tests that, for lack of enough standardization and reliability, have recurrence rates in the order of 33%. Spectroscopic interrogation of these margins is possible during surgery, but algorithms are needed for parametric or dimension reduction processing. One methodology for tumor discrimination based on dimensionality reduction and nonparametric estimation-in particular, Directional Kernel Density Estimation-is proposed and tested on spectral image data from breast samples. Once a hyperspectral image of the tumor has been captured, a surgeon assists by establishing Regions of Interest where tissues are qualitatively differentiable. After proper normalization, Directional KDE is used to estimate the likelihood of every pixel in the image belonging to each specified tissue class. This information is enough to yield, in almost real time and with 98% accuracy, results that coincide with those provided by histological H&E validation performed after the surgery.