Computer-Aided Cancer Diagnosis and Grading via Sparse Directional Image Representation

Prostate cancer and breast cancer are the second cause of death among cancers in males and females, respectively. If not diagnosed, prostate and breast cancers can spread and metastasize to other organs and bones and make it impossible for treatment. Hence, early diagnosis of cancer is vital for patient survival. Histopathological evaluation of the tissue is used for cancer diagnosis. The tissue is taken during biopsies and stained using hematoxylin and eosin (H&E) stain. Then a pathologist looks for abnormal changes in the tissue to diagnose and grade the cancer. This process can be time-consuming and subjective. A reliable and repetitive automatic cancer diagnosis method can greatly reduce the time while producing more reliable results. The scope of this dissertation is developing computer vision and machine learning algorithms for automatic cancer diagnosis and grading methods with accuracy acceptable by the expert pathologists.;Automatic image classication relies on feature representation methods. In this dissertation we developed methods utilizing sparse directional multiscale transforms{ specically shearlet transform{for medical image analysis. We particularly designed theses computer visions-based algorithms and methods to work with H&E images and MRI images. Traditional signal processing methods (e.g. Fourier transform, wavelet transform, etc.) are not suitable for detecting carcinoma cells due to their lack of directional sensitivity. However, shearlet transform has inherent directional sensitivity and multiscale framework that enables it to detect different edges in the tissue images. We developed techniques for extracting holistic and local texture features from the histological and MRI images using histogram and co-occurrence of shearlet coefficients, respectively. Then we combined these features with the color and morphological features using multiple kernel learning (MKL) algorithm and employed support vector machines (SVM) with MKL to classify the medical images.;We further investigated the impact of deep neural networks in representing the medical images for cancer detection. The aforementioned engineered features have a few limitations. They lack generalizability due to being tailored to the specific texture and structure of the tissues. They are time-consuming and expensive and need prepossessing and sometimes it is difficult to extract discriminative features from the images. On the other hand, feature learning techniques use multiple processing layers and learn feature representations directly from the data. To address these issues, we have developed a deep neural network containing multiple layers of convolution, max-pooling, and fully connected layers, trained on the Red, Green, and Blue (RGB) images along with the magnitude and phase of shearlet coefficients. Then we developed a weighted decision fusion deep neural network that assigns weights on the output probabilities and update those weights via backpropagation. The final decision was a weighted sum of the decisions from the RGB, and the magnitude and the phase of shearlet networks. We used the trained networks for classification of benign and malignant H&E images and Gleason grading. Our experimental results show that our proposed methods based on feature engineering and feature learning outperform the state-of-the-art and are even near perfect (100%) for some databases in terms of classication accuracy, sensitivity, specicity, F1 score, and area under the curve (AUC) and hence are promising computer-based methods for cancer diagnosis and grading using images.