Improved Deep Convolutional Neural Networks (DCNN) Approaches for Computer Vision and Bio-Medical Imaging

Date of Award

2018

Degree Name

Ph.D. in Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Advisor: Tarek M. Taha

Abstract

Deep learning is showing tremendous success in variety of application domains and demonstrates state-of-the-art performance over traditional machine learning approaches in the fields of Computer Vision, Speech Recognition, Natural Language Processing (NLP), Bio-Medical imaging, Computational Pathology, and many more. This thesis presents several improved Deep Convolutional Neural Network (DCNN) models including the Inception Recurrent Convolutional Neural Network (IRCNN) and Inception Recurrent Residual Convolutional Neural Networks (IRRCNN), a Recurrent U-Net (RU-Net), a Recurrent Residual U-Net (R2U-Net) model, a R2U-Net regression model, and a Densely Connected Recurrent Network (DCRN). These models are evaluated for classification, segmentation, and detection tasks in computer vision, Bio-medical imaging, and computational pathology applications. There are four key contribution areas in this thesis.The first contribution area is the introduction of two improved DCNN models for classification tasks: IRCNN and IRRCNN, which utilize the power of the Recurrent Convolutional Neural Network (RCNN), the Inception Network, and the Residual Network (ResNet). In addition, we have evaluated the impact of recurrent convolutional layers on DenseNet which is called Densely Connected Recurrent Network (DCRN). The performance of the IRCNN, DCRN, and IRRCNN models was investigated with a set of experiments and computer vision tasks where we used several publicly available datasets including MNIST, CIFAR 10, CIFAR 100, SVHN, CU3D-100, and Tiny ImageNet-200. The experimental results show that IRCNN, DCRN, and IRRCNN provide superior performance compared to the equivalent DCNN based methods including equivalent RCNN, ResNet, Inception V3, DenseNet, and Inception Residual Network (Inception V-4) with the same number of network parameters for different computer vision tasks. The second contribution area is the introduction of two different models including a Recurrent U-Net and Recurrent Residual U-Net models, which are named RU-Net and R2U-Net respectively. The proposed models utilize the power of U-Net, the Residual Network, the RCNN, and U-Net for image segmentation tasks. These proposed architectures have several advantages for segmentation tasks over the existing DL methods. First, a residual unit helps when training deep architectures. Second, feature accumulation with recurrent residual convolutional layers ensures better feature representation for segmentation tasks. Third, it allows us to design better U-Net architecture with the same number of network parameters with better performance for medical image segmentation. The proposed models are tested on three benchmark datasets for blood vessel segmentation in retina images, skin cancer segmentation, and lung lesion segmentation. The experimental results show superior performance on segmentation tasks compared to equivalent models including SegNet, U-Net and Residual U-Net (ResU-Net) in different Bio-medical segmentation tasks.The third contribution area is the introduction of an R2U-Net based regression model which is named University of Dayton Network (UD-Net) and is used for end-to-end detection tasks in digital pathology. To generalize these advanced DCNN models, we have applied classification, segmentation, and detection tasks in Digital Pathology Image Analysis (DPIA) including: microscopic blood cell classification, Breast Cancer Classification (BCC), invasive ductal carcinoma detection, and lymphoma classification, nuclei segmentation, epithelium segmentation, tubule segmentation, lymphocyte detection, and mitosis detection. The experiments have been conducted on different publicly available datasets and evaluated with different performance metrics. The results demonstrate superior performance compared to existing DCNN based methods.The fourth contribution area is the introduction of an image reconstruction technique using Convolutional Sparse Coding (CSC) on IBM's TrueNorth Neuromorphic computing system and the results demonstrate promising sparse reconstructions for two different benchmarks: MNIST and CIFAR-10. In 2016, IBM's release of a deep learning framework for DCNNs called Energy Efficient Deep Neuromorphic Networks (EEDN). EEDN shows promise for delivering high accuracies across different benchmark while consuming very low power using IBM's TrueNorth chip. We have empirically evaluated the performance of different DCNN architectures implemented within the EEDN framework to discover the most efficient way to implement DCNN models for object classification tasks using the TrueNorth system. The results show that for datasets with large numbers of classes, wider networks perform better when compared to deep networks comprised of nearly the same core complexity on IBM's TrueNorth system. In addition, we have proposed an effective quantization approach for Recurrent Neural Networks (RNN): Long Short-Term Memory (SLTM), Gated Recurrent Unit (GRU), and Convolutional LSTM (ConvLSTM). Furthermore, an NP-hard optimization problem called Quadratic Unconstrained Binary Optimization (QUBO) has solved with vanilla RNN on IBM's Neuromorphic computing system.

Keywords

Bioinformatics, Biomedical Research, Computer Science, Electrical Engineering, Pathology, Deep Learning, DCNN, IRRCNN, DCRN, Medical Imaging, Computer Vision, Computational Pathology

Rights Statement

Copyright © 2018, author

Share

COinS