Active contour-based methods are widely popular in the image segmentation field. Basically, they perform a semiautomatic region identification by partitioning the image content mainly into the foreground and background. Nevertheless, the accurate delimitation still remains as an important challenge which usually depends on how close the initial contour is placed to the object of interest (OI). Several applications of active contours require the user interaction to give prior information about the initial position as the first step, which drives the tool substantially dependent on a manual process. This paper describes how to overcome this limitation by including the expertise provided by the training stage of a Convolutional Neural Network (CNN). Despite CNN methods require a large dataset or data augmentation techniques to improve their results, the combined proposal accomplishes a presegmentation task with a reduced number of images to obtain the assumed locations for each OI. These results are used to initialize a multiphase active contour model that follows a level set scheme to lead a smoother multiregion segmentation with less effort. Experiments of this approach are included to compare classic techniques of contour initialization and show the benefits of our proposal.
Superpixel algorithms oversegment an image by grouping pixels with similar local features such as spatial position, gray level intensity, color, and texture. Superpixels provide visually significant regions and avoid a large number of redundant information to reduce dimensionality and complexity for subsequent image processing tasks. However, superpixel algorithms decrease performance in images with high-frequency contrast variations in regions of uniform texture. Moreover, most state-of-the-art methods use only basic pixel information -spatial and color-, getting superpixels with low regularity, boundary smoothness and adherence. The proposed algorithm adds texture information to the common superpixel representation. This information is obtained with the Hermite Transform, which extracts local texture features in terms of Gaussian derivatives. A local iterative clustering with adaptive feature weights generates superpixels preserving boundary adherence, smoothness, regularity, and compactness. A feature adjustment stage is applied to improve algorithm performance. We tested our algorithm on Berkeley Segmentation Dataset and evaluated it with standard superpixel metrics. We also demonstrate the usefulness and adaptability of our proposal in medical image application.
Ultrasound (US) has become one of the most common forms for medical imaging in clinical practice. It is a non-invasive and safe practice that allows obtaining images in real time. It is also a technology with important challenges such as low image quality and high variability (between manufacturers and institutions) . This work aims to apply a fast and accurate deep learning architecture to detect and locate cerebellum in prenatal ultrasound images. Cerebellum biometry is used to estimate fetal age  and cerebellum segmentation could be applied to detect malformation . YOLO (You Only Look Once) is a convolutional neural network (CNN) architecture for detection, classification and location of objects in images . YOLO was innovative because it solved a regression problem to predict the location (coordinates and sizes) of bounding boxes and associated classes. We used 316 ultrasound scans of fetal brains and their respective cerebellar segmentations. From these, 78 images were randomly taken to be treated as test images and the rest were available to feed the trainings. Segmentation masks were converted to numerical descriptions of bounding boxes. To deal with small data set, transfer learning was done by initializing convolutional layers with weights pretrained on Imagenet . We evaluated detection using F1 score and localization using average precision (AP) for 78 test images. Our best AP was 84.8% using 121 divisions or cells per image. Future work will focus on segmentation task assisted by localization.
In this paper we propose a semi-automatic method to segment the fetal cerebellum in ultrasound images. The method is based on an active shape model which includes profiles of Hermite features. In order to fit the shape model we used a PCA of Hermite features. This model was tested on ultrasound images of the fetal brain taken from 20 pregnant women with gestational weeks varying from 18 to 24. Segmentation results compared to manual annotation show a mean Hausdorff distance of 6.85 mm using a conventional active shape model trained with gray profiles, and a mean Hausdorff distance of 5.67 mm using Hermite profiles. We conclude that the Hermite profile model is more robust in segmenting fetal cerebellum in ultrasound images.
Texture is one of the most important elements used by the human visual system (HVS) to distinguish different objects in a scene. Early bio-inspired methods for texture segmentation involve partitioning an image into distinct regions by setting a criterion based on their frequency response and local properties in order to further perform a grouping task. Nevertheless, the correct texture delimitation still remains as an important challenge in image segmentation. The aim of this study is to generate a novel approach to discriminate different textures by comparing internal and external image content in a set of evolving curves. We propose a multiphase formulation with an active contour model applied on the highest energy coefficients generated by the Hermite transform (HT). Local texture features such as scale and orientation are reflected in the HT coefficients which guide the evolution of each curve. This process leads to the enclosure of similar characteristics in a region associated with a level set function. The efficiency of our proposal is evaluated using a variety of synthetic images and real textured scenes.
Periodic variations in patterns within a group of pixels provide important information about the surface of interest and can be used to identify objects or regions. Hence, a proper analysis can be applied to extract particular features according to some specific image properties. Recently, texture analysis using orthogonal polynomials has gained attention since polynomials characterize the pseudo-periodic behavior of textures through the projection of the pattern of interest over a group of kernel functions. However, the maximum polynomial order is often linked to the size of the texture, which implies in many cases, a complex calculation and introduces instability in higher orders leading to computational errors. In this paper, we address this issue and explore a pre-processing stage to compute the optimal size of the window of analysis called “texel.” We propose Haralick-based metrics to find the main oscillation period, such that, it represents the fundamental texture and captures the minimum information, which is sufficient for classification tasks. This procedure avoids the computation of large polynomials and reduces substantially the feature space with small classification errors. Our proposal is also compared against different fixed-size windows. We also show similarities between full-image representations and the ones based on texels in terms of visual structures and feature vectors using two different orthogonal bases: Tchebichef and Hermite polynomials. Finally, we assess the performance of the proposal using well-known texture databases found in the literature.
Medical image analysis has become an important tool for improving medical diagnosis and planning treatments. It involves volume or still image segmentation that plays a critical role in understanding image content by facilitating extraction of the anatomical organ or region-of-interest. It also may help towards the construction of reliable computer-aided diagnosis systems. Specifically, level set methods have emerged as a general framework for image segmentation; such methods are mainly based on gradient information and provide satisfactory results. However, the noise inherent to images and the lack of contrast information between adjacent regions hamper the performance of the algorithms, thus, others proposals have been suggested in the literature. For instance, characterization of regions as statistical parametric models to handle level set evolution. In this paper, we study the influence of texture on a level-set-based segmentation and propose the use of Hermite features that are incorporated into the level set model to improve organ segmentation that may be useful for quantifying left ventricular blood flow. The proposal was also compared against other texture descriptors such as local binary patterns, Image derivatives, and Hounsfield low attenuation values.
In recent years, the use of Magnetic Resonance Imaging (MRI) to detect different brain structures such as
midbrain, white matter, gray matter, corpus callosum, and cerebellum has increased. This fact together with
the evidence that midbrain is associated with Parkinson’s disease has led researchers to consider midbrain
segmentation as an important issue. Nowadays, Active Shape Models (ASM) are widely used in literature for
organ segmentation where the shape is an important discriminant feature. Nevertheless, this approach is based
on the assumption that objects of interest are usually located on strong edges. Such a limitation may lead to a
final shape far from the actual shape model. This paper proposes a novel method based on the combined use
of ASM and Local Binary Patterns for segmenting midbrain. Furthermore, we analyzed several LBP methods
and evaluated their performance. The joint-model considers both global and local statistics to improve final
adjustments. The results showed that our proposal performs substantially better than the ASM algorithm and
provides better segmentation measurements.
This paper describes a segmentation method for time series of 3D cardiac images based on deformable models. The goal
of this work is to extend active shape models (ASM) of
tree-dimensional objects to the problem of 4D (3D + time)
cardiac CT image modeling. The segmentation is achieved by constructing a point distribution model (PDM) that
encodes the spatio-temporal variability of a training set, i.e., the principal modes of variation of the temporal shapes are
computed using some statistical parameters. An active search is used in the segmentation process where an initial
approximation of the spatio-temporal shape is given and the gray level information in the neighborhood of the landmarks
is analyzed. The starting shape is able to deform so as to better fit the data, but in the range allowed by the point
distribution model. Several time series consisting of eleven 3D images of cardiac CT are employed for the method
validation. Results are compared with manual segmentation made by an expert. The proposed application can be used
for clinical evaluation of the left ventricle mechanical function. Likewise, the results can be taken as the first step of
processing for optic flow estimation algorithms.