Development of Segmentation and Dimensionality Reduction Techniques for Ultrasound and Magnetic Resonance Image Analysis

分割和降維技術的開發及在超聲和核磁共振圖像上的應用

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date17 Mar 2020

Abstract

Ultrasound (US) and magnetic resonance (MR) imaging are widely used in clinical diagnosis. They provide doctors with non-invasive means to directly observe multiple organs, and capture multi-angle and high-contrast tissue images. Over the past several decades, it has been demonstrated that the two imaging techniques provide great advantages in detecting and diagnosing a variety of diseases. Specifically, for three diseases commonly diagnosed throughout the world, carotid atherosclerosis, prostate cancer, and ischemic cardiomyopathy, these techniques are commonly used. 3D US has proven to be a reproducible and accurate imaging tool for monitoring the progression (or regression) of atherosclerosis. Multiparametric MRI (mpMRI) has been developed to accurately detect and localize prostate cancer foci. 3D Late Gadolinium Enhancement (LGE) MRI widely serves as a standard clinical technique to visualize and identify myocardial scars.

Despite the success of the two imaging techniques in the diagnosis of these three diseases, several problems have emerged as the scale of US and MRI data has rapidly expanded. Existing ad-hoc biomarkers for carotid atherosclerosis could lose their effectiveness in the evaluation of new and existing dietary treatments if US data cannot be evaluated efficiently and sensitively. Manual annotation of the regions of prostates or left ventricular (LV) scars from imaging data is extremely time-consuming and causes a significant increase in financial expenditure. To address these problems and to explore the application of advanced machine learning techniques to the two imaging techniques, this thesis focuses on three ways to improve medical diagnoses: (i) introducing a dimensionality reduction method to develop more sensitive 3D US-derived biomarkers in the evaluation of dietary treatments for carotid atherosclerosis, (ii) delineating prostate lesions from mpMRI based on dimensionality reduction techniques, (iii) accurately and rapidly segmenting left ventricular scars from 3D LGE MRI.

With the continuing development of therapeutic options for atherosclerosis, image-based biomarkers sensitive to the effect of new interventions are required to be developed for cost-effective clinical evaluation. Although 3D ultrasound measurement of total plaque volume (TPV) showed the efficacy of a high-dose statin, more sensitive biomarkers are needed to establish the efficacy of dietary supplements expected to confer a smaller beneficial effect. This study involved 171 subjects who participated in a one-year placebo-controlled trial evaluating the effect of pomegranate. A framework involving a feature selection technique known as discriminative feature selection (DFS) and a semi-supervised graph-based regression (SSGBR) technique was proposed for sensitive detection of plaque textural changes over the trial. 376 textural features of plaques were extracted from 3D ultrasound images acquired at baseline and a follow-up session. A scalar biomarker for each subject were generated by SSGBR based on prominent textural features selected by DFS. The ability of this biomarker to discriminate pomegranate-receiving from placebo subjects was quantified by the p-values obtained in a Mann-Whitney U test. The discriminative power of SSGBR was compared with global and local dimensionality reduction techniques, including linear discriminant analysis (LDA), maximum margin criterion (MMC) and Laplacian Eigenmap (LE). The result demonstrated that the biomarker generated by SSGBR are more able to detect statically significant changes between subjects treated by pomegranate and placebo. Only SSGBR (p = 4.12×10-6) and normalized LE (p = 0.002) detected a difference between the two groups at the 5% significance level. As compared with ∆TPV, SSGBR reduced the sample size required to establish a significant difference by a factor of 60. The application of this framework will substantially reduce the cost incurred in clinical trials.

The goal of this study is to develop and evaluate an algorithm that provides pixel-accurate lesion delineation from images acquired based on mpMRI. Pixel-wise classification was performed on the reduced space generated by locality alignment discriminant analysis (LADA), a version of linear discriminant analysis (LDA) localized to patches in the feature space. Post-processing procedures, including removal of isolated points identified and filling of holes inside detected regions, were performed to improve delineation accuracy. The segmentation result was evaluated against the lesions manually delineated by four expert observers according to the Prostate Imaging-Reporting and Data System (PI-RADS) detection guideline. The LADA-based classifier (60 ±11 %) achieved a higher sensitivity than the LDA-based classifier (51 ±10 %), thereby demonstrating, for the first time, that higher classification performance was attained on the reduced space generated by LADA than by LDA. Further sensitivity improvement (75 ±14 %) was obtained after post-processing, approaching the sensitivities attained by previous mpMRI lesion delineation studies in which non-clinical T2 maps were available. The proposed algorithm delineated lesions accurately and efficiently from images acquired following the clinical protocol. The development of this framework may potentially accelerate the clinical uses of mpMRI in prostate cancer diagnosis and treatment planning.

While 3D three-dimensional (3D) late gadolinium enhanced (LGE) magnetic resonance (MR) imaging provides good conspicuity of small myocardial lesions with short acquisition time, it poses a challenge for image analysis as a large number of axial images are required to be segmented. In particular, prior segmentation of the myocardium, required for most scar segmentation methods, cannot be completed in a clinically feasible time-frame for 3D LGE MR images. We developed a convolutional neural network (CNN) called deep multi-scale residual U-Net (DMS-ResUNet) to segment myocardial scar from 3D LGE MRI without prior myocardium segmentation. DMS-ResUNet has 51 convolutional levels along the encoder path. The use of components in ResNet has made convergence of a CNN with this depth possible. Optimization was further enhanced by the use of side-outputs of different scales in constructing the loss function. This deep supervision feature, along with the multi-scale inputs, allow DMS-ResUNet to learn from global and local features. DMS-ResUNet was evaluated in five-fold cross-validation experiments involving 34 left ventricular LGE MR images. The results were compared with those generated by the fully convolutional network (FCN), U-Net and conventional thresholding methods, including full-width-at-half-maximum and signal-threshold to reference-mean. The sensitivity attained by DMSResUNet was 9.8% and 8.4% higher than FCN and U-Net, respectively, with the same specificity level. DMS-ResUNet does not require human interaction, and the time required for segmenting each MR volume was 1.7 ± 0.6 s. The accuracy and efficiency afforded by DMS-ResUNet in scar segmentation will make possible future studies involving a large population.

    Research areas

  • Segmentation, Dimensionality Reduction Techniques, Ultrasound, Magnetic Resonance Image