Please use this identifier to cite or link to this item:
Scopus Web of ScienceĀ® Altmetric
Full metadata record
DC FieldValueLanguage
dc.contributor.advisorCarneiro, Gustavo-
dc.contributor.advisorBradley, Andrew P.-
dc.contributor.advisorChin, Tat-Jun-
dc.contributor.advisorLu, Zhi-
dc.contributor.authorDhungel, Neeraj-
dc.description.abstractBreast cancer is considered to be one of the major contemporary problems affecting the lives of thousands of women worldwide. One of the most effective tools in the fight against this disease is early detection based on the manual analysis of X-ray mammograms. This manual process of interpretation of mammograms involves the detection of breast lesions (e.g., masses), the segmentation of lesions boundaries and the classification of lesions based on their shape, appearance and texture features. This manual analysis of breast lesions from mammograms presents large interpretation variability amongst radiologists. This variability can be reduced with the aid of computer aided diagnosis (CAD) systems that can act as a second reader in the analysis of breast lesions. However, for a CAD system to be useful in a clinical setting, it must effectively classify lesions as benign or malignant. Detection, segmentation and classification of breast lesions are the main three steps involved in fully automated CAD systems that can work in the analysis of mammograms. Building a CAD system is difficult because mammograms are marred by low signal to noise ratio for the visualisation of breast lesions. In addition, breast lesions present a large variation in terms of shape, size and appearance. A large number of methods have been applied for building automated CAD systems for both types of lesions, namely mass and micro-calcification, but in this work we focus only on the analysis of masses. The major drawback of current approaches is that they generate a large number of false positives and miss a fair amount of true positive regions during the mass detection stage. Furthermore, mass segmentation is generally based on active contour models and graph-based approaches that rarely capture the large shape and appearance variations of breast masses. Finally, mass classification is generally implemented using sub-optimal hand-crafted features and machine learning classifiers such as support vector machines (SVM), linear discriminant analysis (LDA), artificial neural net (ANN), etc. One major limitation of the majority of existing CAD systems is that most of them require manual intervention to obtain mass candidates for segmentation and classification. This thesis presents a new approach based on recently developed deep learning models to develop a fully automated CAD system for automated detection, segmentation and classification of masses from mammograms. Our proposed solution to the mass detection problem consists of three stages: 1) mass candidate generation using multi-scale deep learning and Gaussian mixture models, 2) false positive reduction with a cascade of deep learning and random forests classifiers, 3) candidate refinement with a local search algorithm based on Bayesian optimisation. Our proposed mas segmentation methods are based on two kinds of structured output learning methods, namely: 1) structured support vector machine for parameter estimation and graph cut for inferring the segmentation labels, and 2) truncated fitting for parameter learning and tree re-weighted belief propagation for inference. The resulting segmentation is then refined using an active contour model. Our proposed mass classification deep learning method is modelled with a two-step training procedure, where the first step is based on a pre-training stage that estimates a large set of hand-crafted features, which is followed by a fine-tuning step that learns a classifier (that classifies masses into benign and malignant). Finally, we integrate our mass detection, mass segmentation and mass classification methods into a fully automated CAD system for the analysis of masses in mammograms. We validate our methodology on two publicly available datasets (INbreast and DDSM-BCRP) using different performance measures such as average Dice index for segmentation, free receiver operating curve (FROC) and average precision curve for detection, receiver operating curve (ROC), area under curve (AUC) and accuracy for classification. The experiments show that our methodology for detection, segmentation and classification of breast masses achieves competitive results with respect to the current state-of-the-art techniques in terms of all performance measures mentioned above.en
dc.subjectdeep learningen
dc.subjectgraphical modelen
dc.subjectResearch by Publication-
dc.titleAutomated detection, segmentation and classification of masses from mammograms using deep learningen
dc.contributor.schoolSchool of Computer Scienceen
dc.provenanceCopyright material removed from digital thesis. See print copy in University of Adelaide Library for full text.en
dc.provenanceThis electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at:
dc.description.dissertationThesis (Ph.D.) (Research by Publication) -- University of Adelaide, School of Computer Science, 2016.en
Appears in Collections:Research Theses

Files in This Item:
File Description SizeFormat 
01front.pdf149.34 kBAdobe PDFView/Open
02whole.pdf13.42 MBAdobe PDFView/Open
  Restricted Access
Library staff access only430.01 kBAdobe PDFView/Open
  Restricted Access
Library staff access only20.11 MBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.