Edwin L. Dove

The fuzzy Hough transform-feature extraction in medical images

Identification of anatomical features is a necessary step for medical image analysis. Automatic methods for feature identification using conventional pattern recognition techniques typically classify an object as a member of a predefined class of objects, but do not attempt to recover the exact or approximate shape of that object. For this reason, such techniques are usually not sufficient to identify the borders of organs when individual geometry varies in local detail, even though the general geometrical shape is similar. The authors present an algorithm that detects features in an image based on approximate geometrical models. The algorithm is based on the traditional and generalized Hough Transforms but includes notions from fuzzy set theory. The authors use the new algorithm to roughly estimate the actual locations of boundaries of an internal organ, and from this estimate, to determine a region of interest around the organ. Based on this rough estimate of the border location, and the derived region of interest, the authors find the final (improved) estimate of the true borders with other (subsequently used) image processing techniques. They present results that demonstrate that the algorithm was successfully used to estimate the approximate location of the chest wall in humans, and of the left ventricular contours of a dog heart obtained from cine-computed tomographic images. The authors use this fuzzy Hough transform algorithm as part of a larger procedure to automatically identify the myocardial contours of the heart. This algorithm may also allow for more rapid image processing and clinical decision making in other medical imaging applications.

Automatic detection of myocardial contours in cine-computed tomographic images

Quantitative evaluation of cardiac function from cardiac images requires the identification of the myocardial walls. This generally requires the clinician to view the image and interactively trace the contours. This method is susceptible to great variability that depends on the experience and knowledge of the particular operator tracing the contours. The particular imaging modality that is used may also add tracing difficulties. Cine-computed tomography (cine-CT) is an imaging modality capable of providing high quality cross-sectional images of the heart. CT images, however, are cluttered, i.e., objects that are not of interest, such as the chest wall, liver, stomach, are also visible in the image. To decrease this variability, investigators have developed computer-assisted or near-automatic techniques for tracing these contours. All of these techniques, however, require some operator intervention to confidently identify myocardial borders. The authors present a new algorithm that automatically finds the heart within the chest, and then proceeds to outline (detect) the myocardial contours. Information at each tomographic slice is used to estimate the contours at the next tomographic slice, thus allowing the algorithm to work in near-apical cross-sectional images where the myocardial borders are often difficult to identify. The algorithm does not require operator input and can be used in a batch mode to process large quantities of data. An evaluation and correction phase is included to allow an operator to view the results and selectively correct portions of contours. The authors tested the algorithm by automatically identifying the myocardial borders of 27 cardiac images obtained from three human subjects and quantitatively comparing these automatically determined borders with those traced by an experienced cardiologist.

Quantitative shape descriptors of left ventricular cine-CT images

In order to assess regional diastolic function of the left ventricle (LV), LV rapid-acquisition computed tomographic (cine-CT) images were used to build finite element models. To quantitatively evaluate the accuracy of the geometric reconstruction technique used in building these models, a new measure of shape similarity is introduced. The results obtained from the new measure were compared and contrasted to the results obtained from traditional shape similarity measures. All of these measures were used to compare the endocardial and epicardial LV contours obtained from video images of the same hearts. The results show that the imaging procedure accurately reproduces shape, and further suggest that the descriptor shows the sensitivity and resolution required to distinguish between images separated by as little as 3 mm.<<ETX>>

A method for automatic edge detection and volume computation of the left ventricle from ultrafast computed tomographic images

RATIONALE AND OBJECTIVES.Detection of endocardial and epicardial borders of the left ventricle (LV) using various imaging modalities is time-consuming and prone to interpretive error. An automatic border detection algorithm is presented that is used with ultrafast computed tomographic images of the heart to compute cavity volumes. METHODS.The basal-level slice is identified, and the algorithm automatically detects the endocardial and epicardial borders of images from the basal to the apical levels. From these, the ventricular areas and chamber volumes are computed. The algorithm uses the Fuzzy Hough Transform, region-growing schemes, and optimal border-detection techniques. The crosssectional areas and the chamber volumes computed with this technique were compared with those from manually traced images using canine hearts in vitro (n=8) and studies in clinical patients (n=27). RESULTS.Though the correlation was good (r=.88), the algorithm overestimated the LV epicardial area by 4.8 ±6.4 cm², though this error was not statistically different from zero (P>.05). There was no difference in endocardial areas (r=.95, >P.05). The algorithm tended to underestimate the end-diastolic volume (r=.94) and the end-systolic volume (r=.94), although these errors were not statistically different from zero (P> .05). The algorithm tended to underestimate the ejection fraction (r=.80), although this error was not statistically different from zero ( P>. 05). CONCLUSIONS.Automatic detection of myocardial borders provides the clinician with a useful tool for calculating chamber volumes and ejection fractions. The algorithm, with the corrections suggested, provides an accurate estimation of areas and volumes. This algorithm may be useful for contour border identification with ultrasound, positron-emission tomography, magnetic resonance imaging, and other imaging modalities in the heart, as well as other structures.

Optimal surface detection in intravascular ultrasound using multi-dimensional graph search

A new algorithm is proposed for the detection of optimal surfaces in multi-dimensional datasets. The algorithm is based on multi-dimensional dynamic programming. Cost functions were derived using a cylindrical object model by resampling the raw image data along perpendiculars to a prior contour. Radial and angular smoothing filters were used together with a radial derivative operator to convert the data into a cost function. The algorithm was applied To IVUS image sequences of explanted peripheral arterial segments. The algorithm was assessed by comparing the arterial wall estimates obtained from tracings, obtained slice by slice, by an expert human. The slope and intercept were: 1.06/spl plusmn/0.05, -57/spl plusmn/255 pixels, respectively. The slope and intercepts were not different from unity and zero, respectively (p>0.2). The algorithm is capable of recovering surfaces in 3D, has fixed memory requirements, and is fast.

Automatic detection of left ventricular endocardium in cardiac images

An algorithm is presented which locates a globally optimal endocardium in cardiac images. The algorithm does not make any distinction between convex and nonconvex parts of the endocardium. More complex cost functions can be used which would allow the use of more a priori information. An interesting aspect of the algorithm is that the two searches are independent, and hence the algorithm could be implemented on a multiprocessor machine. If more operator-defined start points are available, then the number of simultaneous searches can be increased, e.g., if two start points are defined, four simultaneous searchers can be performed.<<ETX>>

Quantitative shape analysis of left ventricular reconstruction

The shape of the endocardial and epicardial borders of an excised dog heart digitized and traced from cine computed tomography (CT) images are compared with those obtained from actual dissected cross sections digitized and traced with a video camera. A measure for comparing these images that identifies specific areas in the image where the shapes do not match is described. The results obtained from this technique are compared with those from three traditional shape analysis measures: the Fourier mean-square error distance, the Hausdorff distance, and the boundary matching descriptor. The comparison shows that the shape of the LV images from cine-CT closely agrees with the actual cross sections.<<ETX>>

Quantitative shape analysis of left ventricular cine-CT images

The shape of the endocardial and epicardial borders of excised dog hearts digitized and traced from cine-CT (computer tomography) images have been quantitatively compared with those obtained from actual dissected cross sections digitized and traced with a video camera. Results from a standard Fourier analysis have been compared with those obtained from the proposed boundary matching descriptor method. Both measures indicate that the LV (left ventricle) shape is not altered by the proposed reconstruction procedure. The Fourier analysis indicates that the epicardial borders are more accurately reproduced than the endocardial borders (p<0.05). Results from the introduced technique show that there is no difference in accuracy between the two borders. Reasons are offered for this disparity. The new shape descriptor allows the identification of specific regions in the borders where the shapes do not match. On the basis of these data, the shape of the LV images from cine-CT closely agree with the shape of the actual cross sections obtained by dissection. Errors in shape reproduction appear to be random and not systematic.<<ETX>>

A Graph Search-based Algorithm For Detection Of Closed Contours In Images

An algorithm to detect closed contours in images is presented. The algorithm involves partitioning the image into two parts and conducting simultaneous graph searches in both parts. The algorithm was used to detect endocardial borders in cineCT images.

144 ENERGY EXPENDITURE OF PRETERM INFANTS AS DETERMINED BY SIMULTANEOUS DIRECT AND INDIRECT CALORIMETRY

We measured the energy expenditure of 15 healthy, growing preterm infants (mean weight 1.55 kg, range 1.21 - 1.74 kg) simultaneously by continuous direct (gradient layer) and indirect (open circuit) calorimetry in two body positions, supine and prone, after consecutive feedings. Each measurement began 1 hour after feeding and continued for 2 hours.The results are shown below as mean (and SD) in watts/kg.The energy expenditure values determined by direct and indirect calorimetry were not significantly different.Energy expenditure was 9% lower in the prone position than in the supine position, whether measured by direct or indirect calorimetry (p<0.02). This observation confirms similar findings by Masterson et al (Pediatrics 1987;80:689) and Glatzl-Hawlik and Simbruner (Pediatr Res 1989;26:523) and indicates that energy conservation can be enhanced by increasing the time that preterm infants are kept in the prone position. We speculate that this effect is explained by reduced exposed surface area for heat loss and by lower physical activity while in the prone position.