Biomedical Image Analysis for Cancer Classification Using Heterogeneous Computing Environment

Abstract

Computer aided diagnosis (CAD) has become a routine procedure and valued source of detection mechanism. It helps effective diagnosing the diseases and treatment of the patient. Life threatening diseases require quick, accurate and easily deployable CAD procedures which can serve as first level identification. One such disease is Oral Squamous Cell Carcinoma (OSCC). OSCC is the most common among oral cancer patients across the world. In this research work machine learning based automatic OSCC classifier called Stratified Squamous Epithelial Biopsy Image Classifier (SSE-BIC) is developed. It categorizes Haemotoxylin and Eosin stained (H&E-stained) microscopic images of squamous epithelial layer in four different classes: Normal, Well differentiated, Moderately differentiated and Poorly differentiated. The developed classifier is further optimized for faster speed of operation maintaining the accuracy and efficiency using heterogeneous computing. The developed classifier employs majority voting scheme. A novel segmentation technique is developed to segment cellular regions from cytoplasm and other micro-organs present in the image of H&E-stained biopsy sample of Stratified Squamous Epithelium (SSE). The segmentation is implemented under Heterogeneous Computing Environment (HCE) where NVIDIA GeForce GTX 1050-Ti is utilized for parallel computation. Latest NVIDIA libraries like CuBLAS and Thrust are used to implement parallel processing using CUDA-C. Visually inspected characteristics of the biopsy specimen by pathologists are mapped to various mathematically computable features. These features include colour features, texture features, morphological features and orientation features. Thus, total number of features computed from each image is 305. To improve the efficiency of the classifier and reduce the computational complexity feature selection is carried out over these set of features. Neighbourhood Component Feature Selection (NCFS) technique is employed to short-list the most relevant and non-redundant features from the original feature set. Feature selection in NCFS is handled by stochastic gradient descent-based feature weight estimator. This technique is independent of the choice of the classifier. A detailed analysis of effectiveness of the feature selection is performed. These selected features are used to train and test the developed classifier. The developed classifier composed of five base classifiers namely Support Vector Machine with polynomial kernel of degree 2 and 3, Decision Tree classifier, Random subspace combined Linear Discriminant classifier and Neighbourhood Component Analysis classifier. These are used in majority voting arrangement. Total 776 images are used to extract the features to train, cross-validate and test the classifier. A detailed performance analysis is carried out with individual feature sets and hybrid feature set with feature selection using individual classifiers as well as developed classifier. Overall accuracy of the developed classifier is 95.56%. The adaptation of parallel processing for the image segmentation in the developed algorithm accelerates the processing speed of the classifier by 13.04X per image. The developed algorithm can be used for first level of automatic screening of the SSE biopsy images for OSCC diagnosis.