COMPUTER-AIDED DIAGNOSIS OF BREAST LESION AND SKIN LESION USING DELINEATION AND CLASSIFICATION APPROACHES

Abstract

Unless people experience a cancer-related disease in their own lives or are close to someone who does, they do not want to think and talk about it. This was most likely due to the fact that, in most cases, cancer was an incurable disease, and its diagnosis indicated a person’s approaching death. However, nowadays, the situation has changed. Over recent decades, cancer disease incidence and survival rates have changed considerably in many countries, mainly due to new prevention strategies, novel treatment approaches and lifestyle changes. Recently, a lot more research in the field of breast cancer with new and effective treatment strategies have been reported. Currently, different methodologies have been utilized for effective and efficient analysis of the lesion in full detail. This thesis is divided into seven chapters. To begin with, Chapter 1 discusses the etymology of breast cancer, its causes and the accompanying image modalities like mammograms, ultrasound scans, magnetic resonance imaging (MRI), numerous challenges for their localization, detection and segmentation, and the motivation for performing better work to overcome the aforementioned challenges. Chapter 2 then elucidates an elaborate literature on the segmentation and classification approaches of the used image modalities. In Chapter 3, a deep ensemble-based mammogram feature extraction followed by an Neural Network (NN)-based classification approach is presented. The deep ensemble contains DL sub-architectures that handle the feature extraction process, producing an optimized feature vector, which are then fed to the NN training tool for classification i Abstract among benign and malignant classes. The proposed approach is validated on publically available dataset and state-of-the-art comparable results are obtained. Chapter 4 presents an in-depth approach for Breast Ultrasound (BUS) lesion seg mentation. The proposed method uses residual connections with depthwise separable convolution to allow gradients to pass through the network directly. The model also contains bi-directional ConvLSTMs (BConvLSTM) units to extract features in both forward and backward states during feature extraction. The acquired dataset is pre processed and augmented to enhance image contrast and increase the number of images. The final lesion contours are delineated for the BUS samples taken from two different sources to analyze the proposed approach’s robustness in obtaining acceptable results. In Chapter 5, we presents a breast mass segmentation approach using DL architecture with group convolution in a channel-wise manner to extract minute lesion details. After segmentation, Conditional Random Field (CRF) is used as a post-processing technique that helps to segment and analyze the delineated mass boundaries. In the process, ROI-based breast mass mammogram patches are used and augmented to increase the number of input images which helps the network to produce accurate segmentation maps efficiently. Chapter 6 then presents a dermoscopic lesion delineation method that incorporates GNin every convolution layer, which benefits the normalization process by not affecting with the batch sizes and controlling the distribution of features. Further, the network contains Attention Gates (AGs) in the decoder section, which implements DL-based knowledge of Convolution Neural Network (CNN) to the dermoscopic RGB images and focus on the small and crucial lesions. The differentiable nature of AGs allow the network to be trained during back-propagation, which means the attention coefficients get better at highlighting relevant regions of interest. For elevating the model’s performance, the Tversky Loss (TL) based on Tversky Index (TI) is used to obtain the most desirable performance for lesion segmentation, which is generally an imbalanced class problem. Finally, Chapter 7 concludes the thesis followed by discussing future scopes for the lesion segmentation and classification schemes.