Feature Selection Using Combinatorial Method in
Automated Classification of Biomedical Images

Junhua Ding
East Carolina University, USA


Classification of biomedical images is important for clinical diagnostic. However, manual classification of the images is tedious and error prone. Machine learning methods have been applied for automated classification of biomedical images, but the effectiveness of the classification was not high enough until deep learning method was invented and adapted to the classification. Although deep learning based classification of some types of biomedical images is highly effective, and a few of the classifiers can beat human experts in the classification accuracy. However, the method is not feasible for other type of biomedical images that lack large amount of high quality training samples. In order to address the issue, we proposed an method to learn features from limited number of training samples using deep learning method. Then the classification is built on traditional machine learning method such as Support Vector Machine (SVM). But the number of learned features using deep learning normally is huge. In order to improve the training performance and reduce the amount of required training samples, we used a combinatorial method to select the learned features. The features are first grouped into several categories, and then t-way combination is used to select feature sets. Each feature set is used for training an SVM classification model using the same training data set. The classification performance of different feature sets and randomly selected feature sets are compared. Our experimental results show combinatorial method is effective for selecting a better performance feature set. In addition, we found the classification performance of the selected learned feature set is also higher than the one of our hand crafted feature set.

About the Speakers

Junhua Ding, an Associate Professor of computer science at East Carolina University (ECU), where he has been working for 11 years. Before he joined ECU, he had worked with lead biomedical companies as a software engineer and a project manager for near 8 years. His primary research focus is on Data Analytics and Engineering, particularly as applied to development of clinical diagnostic systems and biomedical informatics. He has published more than 80 peer-reviewed research papers in research journals and conference proceedings. His research is funded by NSF and many other agencies.