Mitosis detection in breast cancer histological images is a particularly interesting, but very challenging task in image processing. Recent results in word challenge pattern recognition competition, showed that by using complicated algorithms the success rates poorly reach 61%. The mitotic cells during their lifecycle (mitosis process phase) and more precisely during metaphase, anaphase and telophase are taking various, not recognizable forms, compared with the healthy cells. Particularly difficult is also the detection of the cells in these kind of images, since the background of each image is too complex to segment the cells, and in many cases each image is following entire different pattern in different categories. The histological images that we use has stained with haematoxylin and eosin for the proper contrast enhancement. To detect the cells in each image, we normalize the entire database, subtracting haematoxylin and eosin effects. This is done because each image has a different aspect ratio of staining and consequently different color intense of background.

For the completion of this project we are given a database of images, which is used to train and test the system. The training procedure conducted with prior knowing of the coordinates (Ground Truth) of each cell in the image, followed by the degree of confidentiality, either is or isn’t a cancer cell. For the identification of the mitotic cells we select a texture - based characteristic as a feature, extracting a feature vector using LBP (Local Binary Patterns) for each cell, estimating the mean feature vector of all the mitotic and non - mitotic cells. In the training samples we claim that the distribution is a Gaussian sphere consisting of equally spaced circular areas relative to the center of the Gaussian distribution. This is a simple case when the covariance matrices for the two classes are identical but otherwise arbitrary. Geometrically, this corresponds to the situation in which the samples fall in hyperellipsoidal clusters, leading us to the simplified nearest – neighbor metric. 
Later on, for every cell in the Test Set an identical feature extraction procedure is used extracting the feature vector once again. Using this metric as a decision measure, we decide for each test cell, if is mitotic or not, attaching it to the proper class and extracting the final Confusion Matrix.