Development of medical image/video segmentation via deep learning models

Pezhman Pour, Mansoureh (2021) Development of medical image/video segmentation via deep learning models. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) pezhman.pour.mansoureh_phd_15032657.pdf - Submitted Version Download (6MB) | Preview

Abstract

Image segmentation has a critical role in medical diagnosis systems as it is mostly the initial stage, and any error would be propagated in the subsequent analysis. Certain challenges, including Irregular border, low quality of images, small Region of Interest (RoI) and complex structures such as overlapping cells in images impede the improvement of medical image analysis. Deep learning-based algorithms have recently brought superior achievements in computer vision. However, there are limitations to their application in the medical domain including data scarcity, and lack of pretrained models on medical data. This research addresses the issues that hinder the progress of deep learning methods on medical data. Firstly, the effectiveness of transfer learning from a pretrained model with dissimilar data is investigated. The model is improved by integrating feature maps from the frequency domain into the spatial feature maps of Convolutional Neural Network (CNN). Training from scratch and the challenges ahead were explored as well. The proposed model shows higher performance compared to state-of-the-art methods by %2:2 and %17 in Jaccard index for tasks of lesion segmentation and dermoscopic feature segmentation respectively. Furthermore, the proposed model benefits from significant improvement for noisy images without preprocessing stage. Early stopping and drop out layers were considered to tackle the overfitting and network hyper-parameters such as different learning rate, weight initialization, kernel size, stride and normalization techniques were investigated to enhance learning performance. In order to expand the research into video segmentation, specifically left ventricular segmentation, U-net deep architecture was modified. The small RoI and confusion between overlapped organs are big challenges in MRI segmentation. The consistent motion of LV and the continuity of neighbor frames are important features that were used in the proposed architecture. High level features including optical flow and contourlet were used to add temporal information and the RoI module to the Unet model. The proposed model surpassed the results of original Unet model for LV segmentation by a %7 increment in Jaccard index.

Actions (login required)

View Item

Downloads