China
Africa
Explore chapters and articles related to this topic
Head CT Analysis for Intracranial Hemorrhage Segmentation
Published in Kayvan Najarian, Delaram Kahrobaei, Enrique Domínguez, Reza Soroushmehr, Artificial Intelligence in Healthcare and Medicine, 2022
Heming Yao, Negar Farzaneh, Jonathan Gryak, Craig Williamson, Kayvan Najarian, Reza Soroushmehr
Conventional image processing techniques have been applied for hematoma segmentation (Table 7.1). In Soroushmehr et al. (2015), the expectation-maximization was applied on a Gaussian Mixture Model to segment four components, including hematoma regions, normal tissues, white-matter regions, and catheters. After that, post-processing was performed by removing bright objects and outlier edges. In Ray et al. (2018), the thresholding technique is used to find the brain tissue and hematoma region clusters. The intensity distribution of pixels is analyzed to segment the hematoma regions. After that, morphological operations were performed as post-processing to get rid of outliers. In Roy et al. (2015), a two-class dictionary for normal tissue and hematoma regions was built using patches from “atlas” CT scans and corresponding manual hematoma segmentation. For a given new CT scan, patches were modeled as a combination of the “atlas” patches in the built dictionary to generate hematoma segmentation. The proposed algorithm was evaluated on CT scans from 25 patients with TBI, and the algorithm has a median Dice score of 0.85.
Level Set Methods in Segmentation of SDOCT Retinal Images
Published in Ayman El-Baz, Jasjit S. Suri, Level Set Method in Medical Imaging Segmentation, 2019
N Padmasini, R Umamaheswari, Yacin Sikkandar Mohamed, Manavi D Sindal
Because of the intensity inhomogeneity of the difference image due to the hemorrhages, exudates and neovascularisation, it is difficult to use one single Gaussian to model the blood vessels and one single Gaussian to model the background. Therefore, a Gaussian mixture model is used to accurately estimate blood vessels. Automatic Choroidal Layer Segmentation Using Markov Random Field and Level Set Method was carried out by Chuang Wang et al. [97]. The region based term, incorporating the neighbouring pixel information with the single pixel log-likelihood function by using the Markov Random Field, is modeled into the level set method; The Gaussian Mixture Model is constructed and updated at each level set iteration, instead of learned offline from a fixed training set; and Finally, the Gaussian Vector Field method is used to estimate more comprehensive narrowband around the zero level set to increase the speed of the segmentation. The energy functional of this segmentation model is formulated as:
Habitat imaging of tumor evolution by magnetic resonance imaging (MRI)
Published in Ruijiang Li, Lei Xing, Sandy Napel, Daniel L. Rubin, Radiomics and Radiogenomics, 2019
Bruna Victorasso Jardim-Perassi, Gary Martinez, Robert Gillies
Low perfusion regions could indicate a habitat of cells that are adapted to survive in hypoxic and acidic conditions. These regions may be inferred by different mpMRI combinations, for example, as shown in Figure 7.1, where a habitat map was created by clustering T2 map, T2* map, diffusion-weighted imaging (DWI), and DCE-MRI data. Application of pixel-by-pixel clustering, in a multidimensional space, leverages the information content that is present in the individual pulse-sequence images. These clustering algorithms reveal the complexity of spatial heterogeneity, and the co-registration of MRI-derived habitat maps with histology is crucial to confirm the cellular phenotypes. In the context of intratumoral heterogeneity, clustering algorithms that allow the mixtures of clusters (habitats), such as Gaussian mixture model, are generally more informative than hard-clustering algorithms (Figure 7.2). Figures 7.1 and 7.2 show four distinct habitats detected by mpMRI, which were spatially corroborated by habitats identified histologically in a pre-clinical model of breast cancer. These maps identified both low and high ADC values in regions of necrosis.
Common Audiological Functional Parameters (CAFPAs) for single patient cases: deriving statistical models from an expert-labelled data set
Published in International Journal of Audiology, 2020
Mareike Buhl, Anna Warzybok, Marc René Schädler, Omid Majdani, Birger Kollmeier
Data from patients assigned to the same category were used to determine training distributions for the categories. The training distributions need to be probability density functions for the respective measurement outcome (e.g. audiogram at a certain frequency) or CAFPA value associated with each category to provide a normalised probabilistic interpretation of the underlying data. The simplest, straightforward assumption is to use a normal distribution described by mean μ and standard deviation σ (GMM1). Additionally, a distribution with two normal components (two parameters for mean μ and standard deviation σ, respectively, and weights of components that sum to 1) was considered possible to account for possible asymmetries in the data (Gaussian mixture model, GMM2). For CAFPAs, also a beta distribution was considered to account for strong asymmetries that are expected due to the interval boundaries [0, 1] of the CAFPAs, and the possible shape of the beta distribution where the highest value is at the interval boundary (Figure 3(c)). The beta distribution shows different shapes dependent on their parameters a and b. μ and σ of the beta distribution can be calculated according to Equations (1) and (2), respectively. Figure 3 shows typical shapes of the three considered distributions.
Power of proteomics and progress in precision medicine – 13th central and eastern European proteomic conference (CEEPC), Ustroń, Poland
Published in Expert Review of Proteomics, 2020
Suresh Jivan Gadher, Piotr Widlak, Hana Kovarova
The ‘New Concepts and Methods’ session was added to the program to highlight the advances in the ‘enabling technologies’ which continue to facilitate advances in Precision Medicine. This session commenced with Joanna Polańska (the Silesian University of Technology, Data Mining Division, Gliwice, Poland) who presented, ‘Machine Learning in the processing of proteomics data.’ Mass Spectrometry Imaging (MSI) is a powerful tool in Proteomics and enables untargeted investigations into the spatial distribution of molecular species in a variety of biological specimens but also brings a tremendous amount of data that require dedicated algorithms for signal analysis. Joanna has developed the comprehensive MSI data processing pipeline consisting of the Gaussian Mixture Model-based feature extraction, ‘intelligent stepwise divisive k-means clustering algorithm’ for tissue heterogeneity modeling and Monte Carlo procedure (MI) for the identification of the most critical features and interactions between features distinguishing healthy tissue from the tumor region. Joanna demonstrated the system performance on the MSI proteomic data collected from the head and neck, thyroid and prostate cancer tissue samples.
Variable diagnostics in model-based clustering through variation partition
Published in Journal of Applied Statistics, 2018
Based on its connection with finite mixture models, model-based clustering enjoys appealing interpretability and promising potentials. If all components are normally distributed, the mixture is called a Gaussian mixture model. Gaussian mixture is the most well-documented and widely applied mixture in the model-based clustering framework. For asymmetric or heavy-tailed clusters, more sophisticated mixtures such as skew-normal [3,25] and skew-t [23] mixtures have been introduced. Through the development of novel mixture distributions, model-based clustering has become sufficiently flexible and capable of capturing various cluster shapes. It has applications in numerous scientific fields such as medicine [38], economics [9], and environmental science [17], just to name a few. Despite its prominence, rather little literature in model-based clustering is devoted to diagnostics of obtained model or partition characteristics [34,45].