How to choose and optimize a classifier for your polarimetric imaging data

Mueller polarimetry is a powerful characterization technique for a variety of samples and a promising optical-biopsy tool for early detection of cancer. Recent advances in Mueller imaging devices allow the collection of large ex-vivo and invivo image databases. Although the technique is sensitive to subtle changes in the micro-organization of tissue, the Mueller matrices of such complex media contain intertwined polarimetric effects and are difficult to interpret. To identify the polarimetric signature of a given tissue modification (cancerous or not), machine learning tools are particularly well suited. However, a statistically sound approach is needed to make the most out of these tools and avoid common pitfalls. We present a global statistical framework based on decision theory. It consists of a complete preprocessing and analysis pipeline for polarimetric bioimages. In the analysis stage, we use a loss-risk-based approach to automatically select the optimal classifier among a library of classifiers. The approach allows to determine the subset of polarimetric parameters of interest, to determine the parameters of the classifiers and to assess classifier performance using cross-validation. The proposed framework is illustrated with precancer detection on human ex-vivo cervical samples.