E-Thesis 96 views 302 downloads
Localisation in 3D Images Using Cross-features Correlation Learning / MAJEDALDEIN ALMAHASNEH
Swansea University Author: MAJEDALDEIN ALMAHASNEH
PDF | E-Thesis – open access
Copyright: The author, Majedaldein I. Almahasneh, 2022.Download (71.35MB)
DOI (Published version): 10.23889/SUthesis.59942
Object detection and segmentation have evolved drastically over the past two decades thanks to the continuous advancement in the field of deep learning. Substantial research efforts have been dedicated towards integrating object detection techniques into a wide range of real-world prob-lems. Most ex...
No Tags, Be the first to tag this record!
Object detection and segmentation have evolved drastically over the past two decades thanks to the continuous advancement in the field of deep learning. Substantial research efforts have been dedicated towards integrating object detection techniques into a wide range of real-world prob-lems. Most existing methods take advantage of the successful application and representational ability of convolutional neural networks (CNNs). Generally, these methods target mainstream applications that are typically based on 2D imaging scenarios. Additionally, driven by the strong correlation between the quality of the feature embedding and the performance in CNNs, most works focus on design characteristics of CNNs, e.g., depth and width, to enhance their modelling capacity and discriminative ability. Limited research was directed towards exploiting feature-level dependencies, which can be feasibly used to enhance the performance of CNNs. More-over, directly adopting such approaches into more complex imaging domains that target data of higher dimensions (e.g., 3D multi-modal and volumetric images) is not straightforwardly appli-cable due to the different nature and complexity of the problem. In this thesis, we explore the possibility of incorporating feature-level correspondence and correlations into object detection and segmentation contexts that target the localisation of 3D objects from 3D multi-modal and volumetric image data. Accordingly, we first explore the detection problem of 3D solar active regions in multi-spectral solar imagery where different imaging bands correspond to different 2D layers (altitudes) in the 3D solar atmosphere.We propose a joint analysis approach in which information from different imaging bands is first individually analysed using band-specific network branches to extract inter-band features that are then dynamically cross-integrated and jointly analysed to investigate spatial correspon-dence and co-dependencies between the different bands. The aggregated embeddings are further analysed using band-specific detection network branches to predict separate sets of results (one for each band). Throughout our study, we evaluate different types of feature fusion, using convo-lutional embeddings of different semantic levels, as well as the impact of using different numbers of image bands inputs to perform the joint analysis. We test the proposed approach over different multi-modal datasets (multi-modal solar images and brain MRI) and applications. The proposed joint analysis based framework consistently improves the CNN’s performance when detecting target regions in contrast to single band based baseline methods.We then generalise our cross-band joint analysis detection scheme into the 3D segmentation problem using multi-modal images. We adopt the joint analysis principles into a segmentation framework where cross-band information is dynamically analysed and cross-integrated at vari-ous semantic levels. The proposed segmentation network also takes advantage of band-specific skip connections to maximise the inter-band information and assist the network in capturing fine details using embeddings of different spatial scales. Furthermore, a recursive training strat-egy, based on weak labels (e.g., bounding boxes), is proposed to overcome the difficulty of producing dense labels to train the segmentation network. We evaluate the proposed segmen-tation approach using different feature fusion approaches, over different datasets (multi-modal solar images, brain MRI, and cloud satellite imagery), and using different levels of supervisions. Promising results were achieved and demonstrate an improved performance in contrast to single band based analysis and state-of-the-art segmentation methods.Additionally, we investigate the possibility of explicitly modelling objective driven feature-level correlations, in a localised manner, within 3D medical imaging scenarios (3D CT pul-monary imaging) to enhance the effectiveness of the feature extraction process in CNNs and subsequently the detection performance. Particularly, we present a framework to perform the 3D detection of pulmonary nodules as an ensemble of two stages, candidate proposal and a false positive reduction. We propose a 3D channel attention block in which cross-channel informa-tion is incorporated to infer channel-wise feature importance with respect to the target objective. Unlike common attention approaches that rely on heavy dimensionality reduction and computa-tionally expensive multi-layer perceptron networks, the proposed approach utilises fully convo-lutional networks to allow directly exploiting rich 3D descriptors and performing the attention in an efficient manner. We also propose a fully convolutional 3D spatial attention approach that elevates cross-sectional information to infer spatial attention. We demonstrate the effectiveness of the proposed attention approaches against a number of popular channel and spatial attention mechanisms. Furthermore, for the False positive reduction stage, in addition to attention, we adopt a joint analysis based approach that takes into account the variable nodule morphology by aggregating spatial information from different contextual levels. We also propose a Zoom-in convolutional path that incorporates semantic information of different spatial scales to assist the network in capturing fine details. The proposed detection approach demonstrates considerable gains in performance in contrast to state-of-the-art lung nodule detection methods.We further explore the possibility of incorporating long-range dependencies between arbi-trary positions in the input features using Transformer networks to infer self-attention, in the context of 3D pulmonary nodule detection, in contrast to localised (convolutional based) atten-tion . We present a hybrid 3D detection approach that takes advantage of both, the Transformers ability in modelling global context and correlations and the spatial representational characteris-tics of convolutional neural networks, providing complementary information and subsequently improving the discriminative ability of the detection model. We propose two hybrid Transformer CNN variants where we investigate the impact of exploiting a deeper Transformer design –in which more Transformer layers and trainable parameters are incorporated– is used along with high-level convolutional feature inputs of a single spatial resolution, in contrast to a shallower Transformer design –of less Transformer layers and trainable parameters– while exploiting con-volutional embeddings of different semantic levels and relatively higher resolution.Extensive quantitative and qualitative analyses are presented for the proposed methods in this thesis and demonstrate the feasibility of exploiting feature-level relations, either implicitly or explicitly, in different detection and segmentation problems.
ORCiD identifier: https://orcid.org/0000-0002-5748-1760
Computer Science, Machine Learning, Deep Learning, Object Detection, Object Segmentation, Computer Attention
College of Science