11 research outputs found
Duo chi du ce di zhu dong lun kuo xian he ju bu xiang wei xin xi zai chao sheng ying yong zhong de shi yong
No full text在各种临床应用广泛使用的诊断和治疗工具中，超声成像是其中的一个。与其他成像模式相比，比如计算机断层照相法和磁共振成像，超声波检查法有许多优点:没有辐射风险，设备价格低以及能够实时获取图像。很多超声应用的第一步通常是对感兴趣组织和结构的检测和定位。然而，超声图像存在一些特有的伪影，比如高噪声，低信噪比和灰度不均，这些伪影使得检测任务变得困难。此外，感兴趣区域之间的低对比度也使得这一任务变得更加复杂。在这篇论文里，我们深入研究这些图像伪影并提出新的方法来促进临床中的超声应用。首先，我们提出一个多尺度的框架来进行超声图像的分割，这个框架是基于各向异性去噪扩散和测地主动轮廓线的。各向异性去噪扩散是对边缘敏感且专门用于斑点噪声图像的扩散过程，这里它被用来去除超声图像的斑点噪声，我们对每幅输入图像构造一个多尺度的表示方法，随着尺度的增加，噪声被逐渐地消除。之后，多尺度测地主动轮廓线从粗到细渐进地应用到这些尺度来提取物体的边界线。为了避免在低对比度区域出现边界泄漏的情况，我们把不同尺度之间的边界形状相似性结合到传统的测地主动轮廓线模型里作为一个外部约束来指导轮廓线的演化。在合成和临床图像的实验结果证明了我们的方法的优越性。其次，我们提出一个基于相位的方法来检查和测量超声图像里的胎儿腹部轮廓线。我们定义了一个基于局部相位的度量来检测胎儿腹部的边界线，这个度量称为多尺度特征非对称性，它与图像的亮度无关，并且能为图像里特征的重要程度提供一个绝对的测量。为了估计与腹部轮廓线相吻合的椭圆，我们使用一个迭代随机霍夫变换来排除内腹部边界线的影响，从而使得估计的椭圆逐渐收敛到外边界线。在临床超声图像里进行腹部周长测量的实验结果验证了我们的方法与手工的方法有很高的一致性，这也表明我们的方法可以作为一个可靠的工具来进行产科的护理和诊断。Ultrasound imaging is one of the most widely used diagnostic and therapeutic tools for a variety of clinical applications. Compared with other imaging modalities, such as computed tomography and magnetic resonance imaging, ultrasonography has a lot of advantages: free of radiation risk, low cost of acquisition and images are available in real-time. The first step in many ultrasonic applications is usually the detection and localization of interested tissues and structures. However, there are a number of characteristic artifacts in ultrasound images that make the task difficult such as high speckle noise, low signal-to-noise ratio and intensity inhomogeneity. Besides, the low contrast between regions of interest further complicates the processing. In this thesis, we deeply investigate these image artifacts and propose new techniques to facilitate ultrasonic applications in clinic.First, we propose a multiscale framework for ultrasound image segmentation based on speckle reducing anisotropic diffusion(SRAD) and geodesic active contours (GAC). SRAD is an edge-sensitive diffusion tailored for speckled images, and it is adopted here to reduce speckle noise by constructing a multiscale representation for each input image, where the noise is gradually removed as the scale increases. Multiscale geodesic active contours are then applied along the scales in a coarse-to-fine manner to capture the object boundaries progressively. To avoid boundary leakages in low contrast regions, traditional GAC model is modified by incorporating the boundary shape similarity between different scales as an external constraint to guide the contour evolution. Experimental results in both synthetic and clinical images demonstrate the superiority of the proposed approach.Second, we propose a phase-based approach for fetal abdominal contour detection and measurement in ultrasound images.We define a local phase-based measure, called multiscale feature asymmetry (MSFA), from the monogenic signal to detect the boundaries of fetal abdomen. The MSFA measure is intensity invariant and provides an absolute measurement for the significance of features in the image. In order to estimate the ellipse that fits to the abdominal contour, we employ an iterative randomized Hough transform to exclude the interferences of the inner boundaries of the abdomen, after which the estimated ellipse gradually converges to the outer boundaries. Experimental results in clinical ultrasound images validate the high agreement between our approach and manual approach in the measurement of abdominal circumference, indicating that the proposed approach can be used as a reliable tool for obstetric care and diagnosis.Detailed summary in vernacular field only.Detailed summary in vernacular field only.Detailed summary in vernacular field only.Wang, Weiming .Thesis (Ph.D.) Chinese University of Hong Kong, 2014.Includes bibliographical references (leaves 68-84).Abstracts also in Chinese.Wang, Weiming
The effect of milk supplementation on bone mineral density in postmenopausal Chinese women in Malaysia
No full textDietary studies often report low calcium intake amongst post-menopausal Malaysian women and calcium deficiency has been implicated as part of the etiology of age-related bone loss leading to osteoporosis. Therefore, the objective of this study was to examine the effectiveness of high calcium skimmed milk (Anlene Gold™, New Zealand Milk, Wellington, New Zealand) to reduce bone loss in Chinese postmenopausal women. Two hundred subjects aged 55–65 years and who were more than 5 years postmenopausal were randomized to a milk group and control group. The milk group consumed 50 g of high calcium skimmed milk powder daily, which contained 1200 mg calcium (taken as two glasses of milk a day). The control group continued with their usual diet. Using repeated measures ANCOVA, the milk supplement was found to significantly reduce the percentage of bone loss at the total body compared to the control group at 24 months (control −1.04%, milk −0.13%; P<0.001). At the lumbar spine, the percentage of bone loss in the control group was significantly higher (−0.90%) when compared to the milk (−0.13%) supplemented group at 24 months (P<0.05). Similarly, milk supplementation reduced the percentage of bone loss at the femoral neck (control −1.21%, milk 0.51%) (P<0.01) and total hip (control −2.17%, milk −0.50%) (P<0.01). The supplemented group did not experience any significant weight gain over the 24 months. The serum 25-hydroxy vitamin D level improved significantly (P<0.01) from 69.1±16.1 nmol/l at baseline to 86.4±22.0 nmol/l at 24 months in the milk group. In conclusion, ingestion of high calcium skimmed milk was effective in reducing the rate of bone loss at clinically important lumbar spine and hip sites in postmenopausal Chinese women in Malaysia. Supplementing with milk had additional benefits of improving the serum 25-hydroxy vitamin D status of the subjects
Er wei tu xiang he san wei wang ge zhong ceng fen li wen ti de ji shu yan jiu
No full textPh.D.Layer separation problems take the assumption that the input data (e.g., 2D images and 3D meshes) is composed of different layers, and aim to decompose it into different component layers. This decomposition highly benefits many practical applications, but obtaining a perfect separation is very challenging due to its ill-posedness nature. In this thesis, we focus on four specific layer separation problems, including the mesh denoising, ultrasound speckle reduction, structure-preserving image smoothing, and single-image rain streak removal.First, we develop a novel mesh denoising framework based on the main idea denoising each face normal within its piecewise smooth region (isotropic subneighborhood) instead of using the anisotropic neighborhood. Next, we present two works for speckle reduction in ultrasound images. One is based on the low-rank non-local filtering. Another is a novel optimization framework by leveraging the concept of phase congruency and incorporating a feature asymmetry metric into the regularization term of the objective function to effectively distinguish the features and speckle noise. Our fourth work presents another non-local low-rank filtering framework dedicated for structure-preserving image smoothing, where the input 2D image is separated into the prominent structures and small scale texture details. Lastly, we present a novel method for removing rain streaks from a single input image by decomposing it into a rain-free background layer B and a rain-streak layer R. It uses a joint optimization process with three priors to alternate between removing rain-streak details from B and removing non-streak details from R. In all the works, our proposed methods have achieved superior performance against current state-of-the-art methods, and we also show some applications of our solutions on several tasks, such as edge detection, texture enhancement, seam carving and the breast ultrasound image segmentation.在層分離研究問題中，輸入數據（例如2D圖像和3D網格）通常假 定是由不同的層組成，該類問題的研究目標是將輸入數據分解成不同 的組件層。 儘管這種分解技術極大地有益於許多實際應用，然而由 於其自身不適定的特性，得到一個完美的層分離是非常困難的。 在 本論文中，我們重點解決四個具體的層分離問題： 網格去噪，超聲 斑點噪聲去除，結構特徵保留的圖像平滑和單圖像雨滴去除。首先，我們提出了一個基於各項同性鄰域, 而非傳統各向異性鄰 域, 的法向量濾波的新型網格去噪框架。 接下來，我們提出了兩個超 聲圖像中斑點噪聲去除算法。其中一個是基於低秩的非局部濾波框 架。 另一種是基於相位一致性概念的全域優化的方法，該優化框架 通過將特徵不對稱度量納入目標函數的約束項來有效區分特徵和斑點 噪聲。 在第四個工作中，我們提出了一個基於非局部低秩濾波框架 的結構特徵保留的圖像平滑算法，用來將輸入的2D圖像分為顯著的 結構特徵和小尺度的紋理細節。 最後，我們提出了一個將單張輸入 雨滴圖像分解為無雨滴的背景層B和只有雨滴的雨滴層R的方法。 該 方法採用了一個基於三個先驗的聯合優化方法，來交替實現在B層中 去除雨滴，和在R層中去除非雨滴細節的目標。 在這四個具體層分離 問題中，我們提出的方法已經取得了比當前最先進的方法更好的優越 性能，並且我們還展示了我們的方法在一些實際問題中的具體應用， 比如說邊緣檢測，紋理增強，影像細縫裁減和乳房超聲圖像分割.Zhu, Lei.Thesis Ph.D. Chinese University of Hong Kong 2017.Includes bibliographical references (leaves 100-112).Abstracts also in Chinese.Title from PDF title page (viewed on 14, January, 2020).Zhu, Lei
Gao wei you xian shu ju de tong ji xue xi yan jiu
No full textPh.D.With advancing applications comes the need of information extraction from increasingly high-dimensional data, whereas the sample size is rather limited. However, the analysis of such data is very challenging in statistics due to nondeterminacy issues unless additional assumptions are introduced. To tackle this problem, this thesis develops four structured sparsity constrained statistical learning methods for sample-limited high-dimensional data.We first propose an online subspace learning method to address the challenging image alignment problem. We decompose the aligned image gradient orientation (IGO) of a new image into a sparse error and a linear combination of IGO-PCA basis learned from previously well-aligned IGO. The IGO-PCA basis is further automatically updated to learn the low-dimensional projection. Extensive experiments are conducted to validate the efficiency of our method.Secondly, we investigate a robust matrix classification method called “RSMM” to account for the intra-sample outliers within the high-dimensional data. The idea of RSMM is to simultaneously perform feature recovery and train the classifier based on the clean features. For feature recovery, RSMM assumes that each matrix data can be decomposed into a latent low-rank clean matrix plus a sparse noise matrix. We formulate the RSMM in a unified framework and present an effective solver based on alternating direction method of multipliers (ADMM).Thirdly, we study a novel matrix classification method called “SSMM” to simultaneously select useful features and leverage the structural information for classification improvement. SSMM is defined as a hinge loss plus a combination of nuclear norm and ℓ₁ norm of the regression matrix. It not only captures the intrinsic topological structural information but also simplifies the model by reducing the dimension to a subset of discriminative features for prediction.Lastly, to handle high-dimensional data in tensor form, we propose a regularized multi-task learning method for joint feature selection and tensor classification. For feature selection, we employ the Fisher discriminant criterion to both select discriminative features and control the within-class non-stationarity. For classification, we take both shared and task-specific structural information into consideration. We decompose the regression tensor for each task into a linear combination of a shared tensor and a task-specific tensor, and propose a composite tensor norm. Experimental results on real electroencephalography (EEG) datasets demonstrate the superiority of our method.隨著應用的不斷發展，我們越來越需要從樣本量有限的高維數據提取信息。由於非確定性問題，這些數據的分析非常具有挑戰性，除非引入其他的假設。針對這個難題，本文提出了四種結構化的稀疏約束統計學習方法，用於樣本量有限的高維數據研究。我們首先提出了一個在線子空間學習方法來解決具有挑戰性的圖像對齊問題。我們將新圖像的對齊圖像梯度方向分解為稀疏誤差和從先前對齊良好的圖像梯度方向學習到的主成分基的線性組合。然後主成分基將自動更新進一步學習低維的投 影。大量的實驗驗證了我們方法的有效性。其次，我們研究了一個稱為“RSMM”的魯棒矩陣分類方法來解決高維數據的樣本內噪聲問題。RSMM的思想是執行特徵恢復的同時在乾淨矩陣數據上訓練分類器。對於特徵恢復，RSMM假設每個矩陣數據可以被分解為一個潛在的低秩矩陣加上一個稀疏噪聲矩陣。我們將RSMM構建於統一的框架并提出了一種有效的基於交替方向乘子法的優化算法。第三，我們研究一個命名為“SSMM”的矩陣分類方法，以同時選擇有用的特徵并利用結構信息來提高分類性能。SSMM被定義為合頁損失加上回歸矩陣的核範數和ℓ₁範數的組合。它不僅能提取內在的拓撲結構信息，還通過只考慮對預測有判 斷力的特徵子集來壓縮維度，達到模型簡化。最後，為了處理張量形式的高維數據，我們提出了一種用於聯合特徵選擇和張量分類的正則化多任務學習方法。對於特徵選擇，我們使用線性判別準則來選擇具有區分性的特徵和控制數據內部的非平穩性。對於分類，我們同時考慮多個任務間共享的和特定任務的結構信息。我們將每個任務的回歸張量分解為共享張量和特定任務張量的線性組合，並提出了一個複合 張量範數。高維腦電信號數據的實驗結果證明我們方法的優越性。Zheng, Qingqing.Thesis Ph.D. Chinese University of Hong Kong 2018.Includes bibliographical references (leaves 127-146).Abstracts also in Chinese.Title from PDF title page (viewed on …).Zheng, Qingqing
Deep learning approaches for 3D point cloud analysis and applications
No full textPh.D.3D point clouds are standard outputs of 3D scanning devices and depth sensors. Due to the popularity of various high-quality depth scanning devices, point clouds are gaining more and more attention as a compact representation for 3D data, and as an effective means for processing 3D geometry. Motivated by the far-reaching impact of deep neural networks on various image-related tasks, a number of recent works fall into the new trend of data-driven point cloud analysis and applications. In this thesis, we present our efforts in developing deep learning approaches for 3D point cloud analysis and applications.First, considering that raw point clouds scanned from LiDAR sensors are often sparse, noisy, and non-uniform, we propose the first deep point cloud upsampling network, namely PU-Net. We formulate the network to learn multi-level point-wise features and then conduct the feature expansion via a multi-branch convolution unit. We next employ the regression layers to reconstruct the upsampled point set from the expanded features.Further, to address the severe downsampling problem near the sharp edges, we extend PU-Net by formulating an edge aware consolidation network, namely EC-Net. Different from PU-Net, EC-Net can arrange more points deliberately along the sharp edges while upsampling points. To do so, we formulate a novel edge-aware joint loss function to drive the network to simultaneously recover point coordinates and identify edge points. Compared with state-of-the-arts, EC-Net can enable more accurate surface reconstructions with sharp edges.Third, we present an unsupervised deep learning approach to detect distinctive regions on 3D shapes. To ease the difficulty of shape analysis, we sample point sets from 3D shapes as network inputs and formulate a deep neural network to predict the per-point distinctiveness by addressing an unsupervised shape clustering task. To drive the network learning in an unsupervised manner, we design a novel clustering-based nonparametric softmax classifier, as well as an adapted contrastive loss. We validate the effectiveness of our approach via extensive experiments, and also present its several typical applications.Lastly, we focus on addressing the rotation-variant issue in 3D point cloud processing, which is a common drawback in existing deep point networks. We design a low-level rotation-invariant representation to replace point coordinates as network inputs. To encode both local and global structures, we further design a deep neural network with region relation convolution to embed these low-level representations into high-level features. Experiments on multiple point cloud analysis tasks confirm the effectiveness and superiority of our method against state-of-the-arts.To sum up, we systematically introduce four different deep learning approaches for 3D point cloud analysis and applications. Though the progress of deep point analysis is significant within the past few years, many important questions are still open, such as the analysis on large-scale point sets, few-shot and unsupervised learning, the application to robot system, etc.三維點雲是三維掃描儀和深度相機的標準輸出。隨著各種各樣高質量掃描儀的普及，點雲越來越受到人們的重視。眾所周知，深度神經網絡對各種圖像分析任務都發揮了巨大的作用。受此驅動，近年來形成了一種用深度神經網絡進行點雲分析和應用的新趨勢。本論文介紹了一系列相關研究成果。首先，由於激光雷達傳感器產生的點雲通常是稀疏的、有噪音的，以及分佈不均勻，我們提出了第一個深度點雲上採樣網絡，並命名為PU-Net。其核心思想是學習每個點的多層次特征，而後通過一個多分支卷積單元對其進行隱式的特征空間擴增。最後通過擬合回歸，將擴增後的特征重構為點雲。進一步地，為了解決銳利邊緣上存在的嚴重稀疏問題，本文將PU-Net 擴展為一個具有邊緣感知能力的點雲上採樣網絡，即EC-Net。與PU-Net 不同，EC-Net 可以在擴增點數的同時，在銳利邊緣上生成更多的點。為此，本文提出了一個具有邊緣感知能力的聯合損失函數。該函數可以驅動網絡在恢復三維坐標點的同時檢測邊緣上的點。因此, 從EC-Net 擴增後的點進行三維網格重建可以更好的保留邊緣特征。第三，本文提出了一種全新的無監督深度學習方法來檢測三維物體上的顯著性區域。為了降低三維物體分析難度，本文從三維物體上採樣點雲作為網絡輸入，建立了一個無監督、基於聚類的深度網絡，該網絡可以為每個點預測其顯著性程度。為了達到無監督的訓練，本文設計了一個基於聚類的非參分類器，以及一個自適應的對比損失函數來聯合監督網絡的訓練。本文通過大量的實驗驗證了方法的有效性，並展示了顯著性區域的三個典型應用。最後，本文介紹了一個解決點雲旋轉不變性問題的工作。本文設計了一種旋轉不變的表征來代替常用的三維笛卡爾坐標作為網絡輸入。為了更好的提取局部和全局特征，本文設計了一個區域關係卷積模塊用來提取高層次特征。本文通過多種點雲分析任務驗證了方法的有效性和優越性。綜上，本文系統的介紹了四個不同的深度學習技術，用以進行三維點雲的分析和應用。雖然利用深度學習技術處理點雲在過去幾年取得了巨大進展，該領域仍然存在很多值得進一步研究的問題，比如，如何對大場景點雲進行分析，如何進行少樣本甚至無監督學習，如何與智能機器人系統進行交互等等。Li, Xianzhi.Thesis Ph.D. Chinese University of Hong Kong 2020.Includes bibliographical references (leaves 129-150).Abstracts also in Chinese.Title from PDF title page (viewed on 10, September 2021)
Maximizing the data utilization efficiency in medical imaging diagnosis: from full supervision to weak supervision
No full textPh.D.The annotated data provided by experienced experts is at the core of AI-powered automatic medical imaging diagnosis systems. It is expensive and hard to obtain on a large scale, particularly in the field of medical imaging, where only domain experts, i.e., radiologists, can provide reliable annotations. In this thesis, the effective usage of the limited data is called data utilization efficiency}. To improve the diagnostic performance of medical imaging diagnosis systems, maximizing the data utilization efficiency is very important. In this thesis, we study with this main goal from full supervision to weak supervision, to address some common challenging problems in the medical image diagnosis field.The computer-aided diagnosis (CAD) has been a long-standing topic in medical imaging computing, including automatic disease diagnosis and abnormal region segmentation from various medical images, e.g., multi-modality MRI, CT scans, dermoscopy images, retinal fundus images, etc. These computer-aided diagnosis techniques show the interpretation of medical images and quantitative measurements, which can assist doctors in determining a more accurate diagnosis or treatment planning. With the availability of the massive amount of data and annotations, deep learning has become a de facto standard approach in a variety of medical image diagnosis applications.In the first part, we tackle typical and challenging problems in CAD under full supervision. Toward a more precise diagnosis, we propose several methods to better utilize the data to extract discriminative features for recognition. First, we propose a 3D multi-scale fully convolutional network (MsFCN) with random modality voxel dropout (RMVD) for automatic intervertebral disc (IVD) localization and segmentation. Our method incorporates multiple scales of IVD in the network and alleviates the co-adaptation issue in multi-modality images, thus contributes to better performance. Second, 3D FCNs would suffer from large computational inefficiency, especially segmenting small tumors from a large amount of 3D medical images, e.g., CT scans. To address this problem, we present a novel and effective network, i.e., H-DenseUNet, for liver and tumor segmentation. Third, the convolutional operation is not rotation equivariant, which adds additional complexity for the network to learn rotation equivariance for the segmentation tasks. In this regard, we propose a deeply supervised rotation equivariant network for skin lesion segmentation. Forth, diseases often occur with complications. For example, diabetic macular edema (DME) is a complication associated with diabetic retinopathy (DR). To better utilize the relationship between these two diseases for accurate prediction, we propose a novel cross-disease attention network (CAN) to jointly grade DR and DME by exploring the individual diseases and also the internal relationship between these diseases.Since the labeled data is hard to obtain, semi-supervised and meta-learning approaches are attracting more attention in medical imaging computing. In the second part, we focus on medical image analysis under weak supervision}. To effectively utilize the unlabeled data, we present a novel transformation consistent self-ensembling model (TCSM) for semi-supervised medical image segmentation, where the network is optimized by the weighted combination of a common supervised loss for labeled inputs only and a regularization loss for both labeled and unlabeled data. Unlike common diseases that have a large amount of available labeled data, rare diseases have extremely low-data regimes, which is very challenging to train a network, and so far, catches very little attention. In this regard, we propose a difficulty-aware meta-learning method to address rare disease classification and demonstrate its capability to classify dermoscopy images.由經驗豐富的專業醫師提供的數據與註釋是人工智能自動醫學影像診斷系統的核心。然而，收集大規模的數據與註釋是非常困難的，特別是在醫學影像領域裏，只有專家可以進行可靠的標註。因此，提高數據利用效率對於提高醫學影像診斷系統的準確率非常重要。在本文中，我們沿著這一研究目標, 從全面監督到弱監督, 以解決醫學影像診斷領域中一些常見的挑戰性問題。計算機輔助診斷（CAD）一直是醫學影像計算中的一個長期課題，包括各種醫學影像的自動疾病診斷和異常區域分割，例如，多模態MRI，CT掃描，皮膚鏡檢查圖像，視網膜眼底圖像等。這些輔助診斷技術解釋了醫學影像，並且提供了定量測量，可以幫助醫生確定更準確地診斷或治療計劃。隨著大量數據和註釋的可用性，深度學習已成為各種醫學影像診斷應用中的標準方法。在第一部分中，我們解決在全監督學習下的計算機輔助診斷的一些具有挑戰性的問題。為了更精確地診斷，我們提出了幾種方法來更好地利用特有的影像數據，從而提取識別特徵並做出診斷。首先，我們提出了一種具有隨機模態體素去除（RMVD）的3D多尺度全卷積神經網絡（MsFCN），用於自動椎间盘定位和分割。我們的方法在網絡中結合了多尺度椎间盘，並減輕了多模態圖像中的過擬合問題，從而有助於提高性能。其次，3DFCN存在計算效率低的問題，特別是在大量3D CT裏面檢測小腫瘤的情況。為了解決這個問題，我們提出了一種新穎有效的網絡，即H-DenseUNet，用於肝臟和腫瘤分割。第三，卷積運算不是旋轉等變的，這增加了網絡的額外複雜性以學習分割任務的旋轉等效性。在這方面，我們提出了一種深度監督的旋轉等效網絡，用於皮膚病變分割。第四，疾病經常伴隨並發症發生。例如，糖尿病性黃斑水腫（DME）是與糖尿病性視網膜病變（DR）相關的並發症。為了更好地利用這兩種疾病之間的關係進行準確預測，我們提出了一種新的跨疾病注意網絡（CAN），通過探索個體疾病以及這些疾病之間的內在關係來聯合評估DR和DME。由於標記數據在實際情況中很難獲取，因此半監督和元學習方法在醫學影像計算中引起了更多關注。在第二部分中，我們將重點放在弱監督下的醫學圖像分析。為了有效地利用未標記的數據，我們提出了一種新的變換一致集成模型（TCSM），用於半監督醫學圖像分割，其中網絡通過共同監督損失函數和正則化損失函數的加權組合來優化。與具有大量可用標記數據的常見疾病不同，罕見疾病具有極低數據量，這對於訓練網絡非常具有挑戰性，並且到目前為止，很少受到關注。在這方面，我們提出了一種難度導向的元學習方法來解決罕見的疾病分類並展示其對罕見皮膚鏡圖像進行分類的能力。Li, Xiaomeng.Thesis Ph.D. Chinese University of Hong Kong 2019.Includes bibliographical references (leaves 126-151).Abstracts also in Chinese.Title from PDF title page (viewed on 27, November, 2020)
Mian xiang chao sheng ying xiang de shen du xue xi fang fa: cong ke xing xing dao lu bang xing
No full textPh.D.Ultrasound is a dominant and unique imaging modality in several important clinic scenarios, like surgical planning and prenatal examinations. Segmenting interest objects, localizing landmarks and measuring biometrics are frequently requested during ultrasound scanning. However, subject to the user dependency, manual analyses often present low reproducibility and high discrepancy. Automatic ultrasound image analyses are highly desired to improve the examination quality and extend the ultrasound to its full use in more applications.The great resurgence of deep learning brings breakthroughs for medical image analysis. Whereas, the challenges in ultrasound images, including the non-standard acquisition, poor image quality, varying appearance shift and large volume, still hinder the progress and robustness of deep learning solutions.In this thesis, we focus on improving the feasibility and robustness of deep learning for ultrasound image analysis. Our contributions are as follows. For application part: (a) we proposed the first deep learning method for prostate segmentation in 2D ultrasound images and achieved state-of-the-art performance. (b) we devised the first fully automated method to simultaneously segment fetus, gestational sac and placenta in ultrasound volumes. This study provides new opportunities for precise monitoring of fetal growth. (c) we created the first work about 3D fetal pose estimation in the literature, with the desire to build a general navigation map for many advanced studies. Tackling the challenges in ultrasound images, our methodology contributions are four-fold. (a) we proposed a novel sequentiality based method to address the boundary ambiguity for ultrasound segmentation. (b) we proposed a case adaptation strategy to cope with the appearance shift under different ultrasound imaging conditions. (c) we successfully customized the deep reinforcement learning to effectively narrow the search space for plane detection in ultrasound volumes. (d) we optimized the GPU memory management under limited GPU resources and proved our strategy in improving volumetric segmentation with large volume input.The efforts in this thesis are dedicated to exploring new chances of ultrasound imaging and making it robust in clinic, especially building a comprehensive system with high accuracy and reliability for automated prenatal ultrasound examinations.醫學超聲影像在諸多臨床檢查中佔有主導和獨特的地位，例如手術導航和產前篩查。分割感興趣目標，定位關鍵解剖結構點以及測量生物學參數是超聲掃查中最爲頻繁的定量分析方式。然而，由於掃查過程對醫師的嚴重依賴，手動分析的結果可重復性低且存在較大用戶間差異。因此，臨床應用中迫切地需要全自動的超聲影像分析，用以改善檢查的質量，并幫助超聲影像在更多的檢查中發揮效力。深度學習的復興爲自動化醫學影像的分析帶來了重大突破。然而，超聲影像中存在的挑戰，例如非標準化的影像采集、較差的影像質量、影像表觀的變異以及容積數據過大等問題，使得深度學習在超聲影像中的研究進展緩慢且算法魯棒性欠佳。本論文重點著眼于提高深度學習方法在超聲影像分析應用中的可行性與魯棒性。本論文的貢獻如下，在擴展應用方面：(a) 我們首次提出使用深度神經網絡分割二維前列腺超聲影像，并取得了最好效果。(b) 我們首次構建了一套深度學習方法，用於在三維超聲影像中全自動地分割胎兒、妊娠囊以及胎盤。該研究爲精確的產前胎兒生長評估提供了更多可能。(c) 我們首次提出在三維超聲影像中全自動地估計胎兒的三維姿態。我們相信該研究能爲更高階的產前超聲研究提供關鍵的導航。在研究方法方面，我們的貢獻如下：(a) 我們提出了融合使用順序化特徵來克服超聲影像中普遍存在的邊緣歧義性，以優化超聲影像的分割。(b) 我們提出了基於個例的模型調控方法，用於有效地應對不同超聲成像條件下超聲影像内出現的表觀變異。(c) 我們成功地定製了一套基於深度增强學習的通用方案，用於在三維超聲中有效地縮小搜索空間并全自動地定位胎兒標準切面。(d) 我們優化了有限GPU 資源限制下的内存管理策略，并首次證明我們所使用的策略具備通用性，且有利于提高深度神經網絡對大容積超聲影像的分割。本論文擴展了深度學習方法在醫學超聲影像中的應用，并且在增强各方法的魯棒性方面做出了多方面研究。特別的，我們試圖構建一套完善的、具備高精度和高可靠性的系統用於全自動的產前超聲影像掃查。Yang, Xin.Thesis Ph.D. Chinese University of Hong Kong 2019.Includes bibliographical references (leaves 141-162).Abstracts also in Chinese.Title from PDF title page (viewed on 19, January, 2021).Yang, Xin
A summary of the Malaysian Clinical Practice Guidelines on the management of postmenopausal osteoporosis, 2022
No full textObjectives: The aim of these Clinical Practice Guidelines is to provide evidence-based recommendations to assist healthcare providers in the screening, diagnosis and management of patients with postmenopausal osteoporosis (OP). Methods: A list of key clinical questions on the assessment, diagnosis and treatment of OP was formulated. A literature search using the PubMed, Medline, Cochrane Databases of Systematic Reviews, and OVID electronic databases identified all relevant articles on OP based on the key clinical questions, from 2014 onwards, to update from the 2015 edition. The articles were graded using the SIGN50 format. For each statement, studies with the highest level of evidence were used to frame the recommendation. Results: This article summarizes the diagnostic and treatment pathways for postmenopausal OP. Risk stratification of patients with OP encompasses clinical risk factors, bone mineral density measurements and FRAX risk estimates. Non-pharmacological measures including adequate calcium and vitamin D, regular exercise and falls prevention are recommended. Pharmacological measures depend on patients’ fracture risk status. Very high-risk individuals are recommended for treatment with an anabolic agent, if available, followed by an anti-resorptive agent. Alternatively, parenteral anti-resorptive agents can be used. High-risk individuals should be treated with anti-resorptive agents. In low-risk individuals, menopausal hormone replacement or selective estrogen receptor modulators can be used, if indicated. Patients should be assessed regularly to monitor treatment response and treatment adjusted, as appropriate. Conclusions: The pathways for the management of postmenopausal OP in Malaysia have been updated. Incorporation of fracture risk stratification can guide appropriate treatment
<b>Observations of occurrence and daily activity patterns of ungulates in the Endau Rompin Landscape, peninsular Malaysia</b>
No full textCamera trap data was used to study occurrence and daily activity patterns in the Endau Rompin Landscape of peninsular Malaysia during 2011, 2013 and 2015 to estimate Malayan Tiger Panthera tigris jacksoni population densities. By-catch data were also collected for seven ungulate species: Barking Deer Muntiacus muntjak, Bearded Pig Sus barbatus, Wild Boar Sus scrofa, Greater Mousedeer Tragulus napu, Lesser Mousedeer Tragulus kanchil, Malayan Tapir Tapirus indicus and Sambar Deer Rusa unicolor. Of these, Bayesian single-season occupancy analysis suggested that Barking Deer were the most widespread and Mousedeer spp. the least widespread during the study period. Bearded Pig, Malayan Tapir and Wild Boar were recorded in more than half of the camera trap area (Sambar Deer was excluded due to small sample size). Daily activity patterns based on independent captures in 2015 suggest that Barking Deer, Bearded Pig and Wild Boar are mostly diurnal, mousedeer species are crepuscular and Malayan Tapir strongly nocturnal. </jats:p
Nociceptin/orphanin FQ – NOP receptor system: novel genetic and pharmacological tools
No full textThe neuropeptide nociceptin/orphanin FQ (N/OFQ) selectively binds and activates the N/OFQ peptide
(NOP) receptor. In cells expressing the NOP receptor N/OFQ inhibits cAMP accumulation and Ca2+
conductance and stimulates K+ currents. Via these mechanisms N/OFQ regulates several biological functions
in the central nervous system (pain, locomotion, memory, emotional responses, food intake), as well as in the
periphery (airways, cardiovascular, genitourinary and gastrointestinal systems). Several research tools
including knockout mice and NOP selective agonists and antagonists have been developed in the past and
used to investigate the role played by this peptidergic system in pathophysiology and to identify possible
therapeutic indications of NOP receptor ligands. The aim of the present study was to make available to the
scientific community novel genetic and pharmacological tools to speed up the process of target validation of
the NOP receptor.
Knockout rats for the NOP receptor gene (NOP(-/-)) have been recently generated. These animals were used
in the present study to investigate their emotional (open field, elevated plus maze, and forced swimming
test), locomotor (drag and rotarod test), and nociceptive (plantar and formalin test) phenotype in comparison
to NOP(+/+) littermates. The results were in line with previous findings obtained with selective NOP
receptor antagonists in mice and rats and with mouse knockout studies and indicated that the blockage of
N/OFQergic signalling elicits antidepressant and motor stimulant effects.
A detailed pharmacological characterization of novel NOP receptor non peptide ligands has been performed.
The compound GF-4 displayed high affinity and potency at recombinant human NOP receptor associated
with pure antagonist properties. This profile was confirmed in N/OFQ sensitive animal tissues. In vivo GF-4
elicited, similar to other NOP antagonists, beneficial effects in animal models of Parkinson disease. The
NOP non-peptide agonists Ro 65-6570, SCH 221510 and compound 6d were characterized in vitro using a
calcium mobilization assay and electrically stimulated mouse and rat vas deferens tissues. The results of
these studies demonstrated that Ro 65-6570 and SCH-221510 behaved as full agonists showing however
some level of NOP selectivity in rat, but not mouse, tissues. Compound 6d did not display NOP selectivity.
Finally, mixed NOP/MOP receptor agonists were generated. [Dmt1]N/OFQ(1-13)NH2 was selected as the
most potent compound. The mixed NOP/MOP full agonist activity and high affinity of [Dmt1]N/OFQ(1-
13)NH2 was confirmed at human recombinant receptors in receptor and [35S]GTPgS binding studies, at rat
spinal cord receptors in [35S]GTPgS binding experiments, and at guinea pig receptors inhibiting neurogenic
contractions in the ileum. In vivo in the mouse tail withdrawal assay in mice [Dmt1]N/OFQ(1-13)NH2 was
also able to elicit a robust antinociceptive effect being more potent than N/OFQ (by 30 fold) and morphine
(by 3 fold). The antinociceptive properties of spinal [Dmt1]N/OFQ(1-13)NH2 were confirmed in non human
primate studies. Collectively these results demonstrate that [Dmt1]N/OFQ(1-13)NH2 behaves as mixed
NOP/MOP agonist and susbtantiate the suggestion that such mixed ligands are worthy of development as
innovative spinal analgesics
