Combinatorial Methods of Feature Selection for Cell Image Classification

Feature selection is a major task in machine learning for selecting the most impactful features that will result in better accuracy and performance. Experiment-based feature selection is one of the main approaches to this task. However, this approach is not feasible in many cases due to the large number of experiments needed in order to effectively select the optimized feature set. In this paper, we propose the use of an experiment based feature selection guided by combinatorial methods such as t-way coverage. This approach can significantly reduce the required number of experiments in a study while selecting the best feature set to achieve a high level of accuracy. We conducted a case study of the feature selection process for a Support Vector Machine (SVM) classification of biomedical images for cell typing. Three thousand labeled images were used in the experimental SVM classification, and 32 features were evaluated for each image. We considered feature sets of different sizes ranging from 6 to 32, and, for each size, we generated t- way combinations of features for t from 2 to 6. The accuracy of each combinatorial combination was evaluated, which allowed us to select the optimal set of features for use in the classification process. The experiment results show that the proposed approach is an effective way to conduct feature selection and that it can be adapted easily for feature selection in other machine learning algorithms.