Piantadosi, Gabriele (2017) Breast Cancer Analysis in DCE-MRI. [Tesi di dottorato]

Piantadosi - Pattern Recognition in Breast DCE-MRI Automatic Cancer Analysis (PhD Thesis) V4 [POST REVIEW].pdf

Download (6MB) | Preview
[error in script] [error in script]
Item Type: Tesi di dottorato
Resource language: English
Title: Breast Cancer Analysis in DCE-MRI
Piantadosi, Gabrielegabriele.piantadosi@gmail.com
Date: 10 April 2017
Number of Pages: 213
Institution: Università degli Studi di Napoli Federico II
Department: Ingegneria Elettrica e delle Tecnologie dell'Informazione
Dottorato: Information technology and electrical engineering
Ciclo di dottorato: 29
Coordinatore del Corso di dottorato:
Riccio, Danieledaniele.riccio@unina.it
Date: 10 April 2017
Number of Pages: 213
Keywords: DCE-MRI, Breast, Cancer, Tumour, Machine Learning
Settori scientifico-disciplinari del MIUR: Area 09 - Ingegneria industriale e dell'informazione > ING-INF/05 - Sistemi di elaborazione delle informazioni
Area 09 - Ingegneria industriale e dell'informazione > ING-INF/06 - Bioingegneria elettronica e informatica
Date Deposited: 09 May 2017 16:41
Last Modified: 08 Mar 2018 13:34
URI: http://www.fedoa.unina.it/id/eprint/11860
DOI: 10.6093/UNINA/FEDOA/11860

Collection description

Breast cancer is the most common women tumour worldwide, about 2 million new cases diagnosed each year (second most common cancer overall). This disease represents about 12% of all new cancer cases and 25% of all cancers in women. Early detection of breast cancer is one of the key factors in determining the prognosis for women with malignant tumours. The standard diagnostic tool for the detection of breast cancer is x-ray mammography. The disadvantage of this method is its low specificity, especially in the case of radiographically dense breast tissue (young or under-forty women), or in the presence of scars and implants within the breast. Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) has demonstrated a great potential in the screening of high-risk women for breast cancer, in staging newly diagnosed patients and in assessing therapy effects. However, due to the large amount of information, DCE-MRI manual examination is error prone and can hardly be inspected without the use of a Computer-Aided Detection and Diagnosis (CAD) system. Breast imaging analysis is made harder by the dynamical characteristics of soft tissues since any patient movements (such as involuntary due to breathing) may affect the voxel-by-voxel dynamical analysis. Breast DCE-MRI computer-aided analysis needs a pre-processing stage to identify breast parenchyma and reduce motion artefacts. Among the major issues in developing CAD for breast DCE-MRI, there is the detection and classification of lesions according to their aggressiveness. Moreover, it would be convenient to determine those subjects who are likely to not respond to the treatment so that a modification may be applied as soon as possible, relieving them from potentially unnecessary or toxic treatments. In this thesis, an automated CAD system is presented. The proposed CAD aims to support radiologist in lesion detection, diagnosis and therapy assessment after a suitable preprocessing stage. Segmentation of breast parenchyma has been addressed relying on fuzzy binary clustering, breast anatomical priors and morphological refinements. The breast mask extraction module combines three 2D Fuzzy C-Means clustering (executed from the three projection, axial, coronal and transversal) and geometrical breast anatomy characterization. In particular, seven well-defined key-points have been considered in order to accurately segment breast parenchyma from air and chest-wall. To diminish the effects of involuntary movement artefacts, it is usual to apply a motion correction of the DCE-MRI volumes before of any data analysis. However, there is no evidence that a single Motion Correction Technique (MCT) can handle different deformations - small or large, rigid or non-rigid - and different patients or tissues. Therefore, it would be useful to develop a quality index (QI) to evaluate the performance of different MCTs. The existent QI might not be adequate to deal with DCE-MRI data because of the intensity variation due to contrast media. Therefore, in developing a novel QI, the underlying idea is that once DCE-MRI data have been realigned using a specific MCT, the dynamic course of the signal intensity should be as close as possible to physiological models, such as the currently accepted ones (e.g. Tofts-Kermode, Extended Tofts-Kermode, Hayton-Brady, Gamma Capillary Transit Time, etc.). The motion correction module ranks all the MCTs, using the QI, selects the best MCT and applies a correction before of further data analysis. The proposed lesion detection module performs the segmentation of lesions in Regions of Interest (ROIs) by means of classification at a pixel level. It is based on a Support Vector Machine (SVM) trained with dynamic features, extracted from a suitably pre-selected area by using a pixel-based approach. The pre-selection mask strongly improves the final result. The lesion classification module evaluates the malignity of each ROI by means of 3D textural features. The Local Binary Patterns descriptor has been used in the Three Orthogonal Planes (LBP-TOP) configuration. A Random Forest has been used to achieve the final classification into a benignant or malignant lesion. The therapy assessment stage aims to predict the patient primary tumour recurrence to support the physician in the evaluation of the therapy effects and benefits. For each patient which has at least a malignant lesion, the recurrence of the disease has been evaluated by means of a multiple classifiers system. A set of dynamic, textural, clinicopathologic and pharmacokinetic features have been used to assess the probability of recurrence for the lesions. Finally, to improve the usability of the proposed work, we developed a framework for tele-medicine that allows advanced medical image remote analysis in a secure and versatile client-server environment, at a low cost. The benefits of using the proposed framework will be presented in a real-case scenario where OsiriX, a wide-spread medical image analysis software, is allowed to perform advanced remote image processing in a simple manner over a secure channel. The proposed CAD system have been tested on real breast DCE-MRI data for the available protocols. The breast mask extraction stage shows a median segmentation accuracy and Dice similarity index of 98% (+/-0,49) and 93% %(+/-1,48) respectively and 100% of neoplastic lesion coverage. The motion correction module is able to rank the MCTs with an accordance of 74% with a 'reference ranking'. Moreover, by only using 40% of the available volume, the computational load is reduced selecting always the best MCT. The automatic detection maximises the area of correctly detected lesions while minimising the number of false alarms with an accuracy of 99% and the lesions are, then, diagnosed according to their stage with an accuracy of 85%. The therapy assessment module provides a forecasting of the tumour recurrence with an accuracy of 78% and an AUC of 79%. Each module has been evaluated by a leave-one-patient-out approach, and results show a confidence level of 95% (p<0.05). Finally, the proposed remote architecture showed a very low transmission overhead which settles on about 2.5% for the widespread 10\100 Mbps. Security has been achieved using client-server certificates and up-to-date standards.


Downloads per month over past year

Actions (login required)

View Item View Item