Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Joy Elias - Morgan State University
Medical images of diverse modalities are often sources of essential information for research and education in biomedical literature. To enable effective and efficient search of these diverse images, it might be advantageous for a retrieval system to automatically identify the image category at first to filter out irrelevant images based on modality (e.g., X-ray), body part (e.g., chest), and orientation (e.g., PA) at different levels. In existing research, this category information is only limited to imaging modality, which is typically extracted from image caption and is often not correct or present. However, features based on image contents, such as color, texture, and shape might improve classification accuracy of images at different levels (such as, modalities, body parts, orientations, specific observation, etc.) from full-text biomedical journal articles. Although, the classification performance is usually highly dependent on the choice of hand-crafted features and effectiveness of image feature vectors. While many feature descriptors have been proposed over the past years, they can be quite complicated and domain-specific. Automatic feature learning from image data has thus emerged as a different trend recently, to capture the intrinsic image features without manual feature design. In this work, we compared the performances of a traditional support vector machines (SVMs) based classification approach in a combined feature space of color histogram, histogram of oriented gradients (HOG), and local binary pattern (LBP) and a Deep Learning-based classification approach based on a convolutional neural network (CNN, or ConvNet) which avoids the process of feature handcrafting. The classification is performed using Python and associated libraries, such as OpenCV, scikit-learn, scikit-image, keras, etc. Both methods are tested and compared on a dataset of 5,000 biomedical images of 31 manually assigned disjoint global categories at three different levels (e.g., modality, body part, and orientation/specific visual observation). Both methods achieve good performance (above 80% accuracy) when compared to approaches in existing literature. The slightly better performance of the feature-based method indicates the absence of large number of labeled data in the training set for deep learning. In future, we will explore a crowdsourcing based approach to generate ground-truth data set in addition to use different augmentation techniques for Deep Learning. References: T. M. Lehmann, M. O. G¨uld, T. Deselaers, D. Keysers, H. Schubert, K. Spitzer, H. Ney, and B. B.Wein, “Automatic categorization of medical images for content-based retrieval and data mining,” Comput. Med. Imag. Graph., vol. 29, pp. 143–155, 2005 M. S. Simpson, D. You, M. M. Rahman, Z. Xue, S. K. Antani, D. Demner-Fushman, and G. R. Thoma, Literature-based biomedical image classification and retrieval, Comput. Med. Imag. Graph., vol. 99, pp. 3–13, 2015
Not SubmittedFunder Acknowledgement(s): NSF HBCU-UP awarded to Md Mahmudur Rahman PhD, Assistant Professor, Computer Science Department, Morgan State University, Baltimore, Maryland.
Faculty Advisor: Md Rahman, md.rahman@morgan.edu
Role: For this research project, one of my main focuses was fixing lines of code and combining different classifiers in order to increase accuracy of classifying images. In regard to fixing lines of code, one of the main issues was making sure that the code for the project can run in Python 2 and 3. Without combining different classifiers in the lines of code, the accuracy of image classification will not increase. Also, moving this project from machine learning into a deep learning setting was one of my main focuses.