Follicle regions in both IHC and H&E stained microscopic images have to be determined in an automatic manner. In this WP, computationally efficient image segmentation methods will be investigated. Furthermore, microscopic images are large in size and some parts of the image do not contain any clinically significant information, therefore they have to be segmented into subregions before further processing to reduce the overall computation cost. For this reason, we will study covariance based and MRF based image segmentation methods.

Recently, Donoser et al. [19] developed a covariance matrix based image segmentation method for regular images. Covariance and co-difference matrix based segmentation method for microscopic images will be also developed in this WP. As described in WP1, the co-difference method will be computationally more efficient than the covariance method, because the co-difference matrix can be computed without performing any multiplications.

Another probabilistic segmentation approach is based on Markov Random Fields (MRF) [20], which model image properties as a random field whose probability density function is assumed to be Gibbsian. Graph cut techniques can be effectively used to find the global minimum of the energy function representing the smoothness [21,22] and consistency of the image data in MRFs and they are fast enough to be practical [23]. These techniques consider a weighted graph for the energy function with two terminal vertices called the source and sink. A cut is a set of edges such that the terminals are separated in two disjoint sets. In weighted graphs, the cost of the cut is defined to be the sum of weights of the edges crossing the cut. The minimum cut problem is to find the cut with the minimum sum. The basic idea is that the minimum cut minimizes the energy either globally or locally. The minimum cut, in turn, can be computed very efficiently by max-flow algorithms.

In the specific problem of Follicle Lymphoma Classification, five output classes are defined (nuclei, cytoplasm, extracellular material, background and red blood cells).
Graph cuts can quite easily incorporate other information or smoothness constraints, e.g. edges, detected features e.g. to define better energy functions to be minimized, so that improved segmentation results can be obtained. More specifically, parameters, such as color, intensity wavelet coefficients and wavelet packets will be used.

In addition, the use of Graph-cut techniques to higher dimensionality images obtained by registrating IHC and H&E images will be studied. This may significantly improve the segmentation result. A registration module for these images is already available at OSU, however the consortium intends to perform additional research in this field ((e.g. based on SIFT feature extraction and matching [24]). In this WP, all of the methods will be compared with each other and the currently available methods. Final result will be a computationally efficient segmentation algorithm, which will not require supercomputers for segmenting microscopic images in the future.