Integrative sparse modeling and classification of biomedical imaging patterns

Date

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

The analysis and characterization of imaging patterns is a significant research area with several applications to biomedicine (computer-aided diagnosis), remote sensing (urban planning, environmental monitoring), homeland security (face recognition, object recognition, biometrics) social networking, and numerous other domains. In this dissertation we study and develop mathematical methods and algorithms for disease diagnosis and tissue characterization. The central hypothesis is that we can predict the occurrence of diseases with certain level of confidence using supervised learning techniques that we apply to medical imaging datasets that include healthy and diseased subjects that can be used for training. In the first stage of this work we propose to diagnose diseased patterns using texture characteristics that are derived from medical imaging modalities. The texture feature set consists of fractal dimension, local binary patterns, discrete wavelet frames, Gabor filters, discrete Fourier and Cosine Transforms, statistical co-occurrence indices, edge histogram, and Law’s energy maps. Next, we implemented feature selection using correlation-based techniques to reduce the feature dimensionality. In the learning stage we employed bagging methods using fast decision tree learners, Random Forests, Bayes network, or na\"{i}ve Bayes techniques. These techniques are also used for comparisons at the later stages of this work. Next, we develop methods for calculation of sparse representations to classify imaging patterns and we explore the advantages of this technique over traditional texture-based classification. We also introduce integrative sparse classifier systems that utilize structural block decomposition to address difficulties caused by high dimensionality. We propose likelihood functions for classification and decision tuning strategies. These likelihood scores may also be used to determine a type of confidence interval for prediction. The two application domains are osteoporosis diagnosis in radiographs of the calcaneus bone, and breast lesion characterization in mammograms. Both of these applications are very significant for improving public health. Osteoporosis results in deterioration of bone quality and affects the quality of life of aging populations. Timely diagnosis of osteoporosis can effectively predict fracture risk and prevent the disease. Furthermore, breast cancer is one of the leading causes of death among women. Early detection and characterization of breast lesions is important for increasing the life expectancy and quality of health of women. We performed bone osteoporosis classification experiments on the TCB challenge dataset and breast lesion characterization experiments Mammographic Image Analysis Society data set. In TCB there are 87 healthy and 87 osteoporotic subjects in the calcaneus trabecular bone. MIAS includes benign and malignant breast cancer lesions. In both of these two data sets, the scans of healthy and diseased subjects show little or no visual differences, and their density histograms have significant overlap. In the experiments, our method of block-based sparse representation produced the best classification accuracy on these two datasets. We compared the conventional sparse representations classification (SRC) and texture-based methods with our method in a leave-one-out (LOO) cross-validation (CV) framework. The top texture-based classification performances are 67.8% ACC (classification accuracy) and 70.9% AUC (Area Under the Receiver Operating Curve) for bone characterization, and 63.4% ACC and 62.1% AUC for breast lesion characterization. The top performance of our integrative sparse model method by using a decision threshold equal to zero is 100% ACC and AUC for bone characterization by block-based maximum a posteriori sparsity-based (BBMAP-S) decision function, as well as 100% for bone characterization by block-based log likelihood sparsity-based (BBLL-S) decision function, 98.6% ACC and 97.8% AUC for breast lesion characterization by BBMAP-S decision function, and 100% ACC and 100% AUC for breast lesion characterization by BBLL-S decision function for breast lesion characterization. We also used 10-fold and 30-fold cross-validation to evaluate the classification performances of our classification methods. The top rate of accuracy produced by the texture-based method is 66.7% and corresponding AUC is 67.5% for bone characterization using 10-fold cross-validation. Our method using integrative sparse models has obtained the highest ACC for 30-fold cross-validation is 69.33% and 70.2% with BBMAP-S decision function. It also achieved 70.7% ACC and 74.4% AUC with BBLL-S decision function for bone characterization. In 10-fold cross-validation experiments for bone characterization, BBLL-S produced 60.6% ACC and 62.5% AUC. In the breast lesion characterization application, the best performance over all the ROI sizes is 71.2% ACC and 69.8% AUC using texture-based methods and the conventional SRC method reached 55.0% ACC and 51.8% AUC using 10-fold cross-validation. For our system, the top performance is 86.7% ACC and 88.2% AUC for 30-fold experiments and 68.9% ACC and 73.7% AUC for 10-fold experiments. Our results show that ensemble sparse representations of imaging patterns provide very good separation between groups of healthy and diseased subjects in two challenging diagnostic applications.

Description

Keywords

Citation

DOI