Mariusz Bajger

Two graph theory based methods for identifying the pectoral muscle in mammograms

Two image segmentation methods based on graph theory are used in conjunction with active contours to segment the pectoral muscle in screening mammograms. One method is based on adaptive pyramids (AP) and the other is based on minimum spanning trees (MST). The algorithms are tested on a public data set of mammograms and results are compared with previously reported methods. In 80% of the images, the boundary of the segmented regions has average error less than 2mm. In 82 of 84 images, the boundary of the pectoral muscle found by the AP algorithm has average error less than 5mm.

Minimum Spanning Trees and Active Contours for Identification of the Pectoral Muscle in Screening Mammograms

Image segmentation based on minimum spanning trees (MST) is used to identify the pectoral muscle in screening mammograms. The segmentation found using the MST is used to initialise an active contour for finding an anatomically reasonable estimate of the boundary of the pectoral muscle. The error is reported in terms of the number of in-correctly assigned pixels. Out of 83 images, 25 images have error rates less than 5 percent and 56 images have error rates less than 10 percent. The nature of the errors encountered indicates that the accuracy of computer algorithms for this task is approaching its practical limit.

Automatic Mass Segmentation Based on Adaptive Pyramid and Sublevel Set Analysis

A method based on sublevel sets is presented for refining segmentation of screening mammograms. Initial segmentation is provided by an adaptive pyramid (AP) scheme which is viewed as seeding of the final segmentation by sublevel sets. Performance is tested with and without prior anisotropic smoothing and is compared to refinement based on component merging. The combination of anisotropic smoothing, AP segmentation and sublevel refinement is found to outperform other combinations.

Mammographic Mass Detection with Statistical Region Merging

An automatic method for detection of mammographic masses is presented which utilizes statistical region merging for segmentation (SRM) and linear discriminant analysis (LDA) for classification. The performance of the scheme was evaluated on 36 images selected from the local database of mammograms and on 48 images taken from the Digital Database for Screening Mammography (DDSM). The Az value (area under the ROC curve) for classifying each region was 0.90 for the local dataset and 0.96 for the images from DDSM. Results indicate that SRM segmentation can form part of an robust and efficient basis for analysis of mammograms.

Multi-organ segmentation of CT images using statistical region merging

Segmentation is one of the key steps in the process of developing anatomical models for calculation of safe medical dose of radiation for children. This study explores the potential of the Statistical Region Merging segmentation technique for tissue segmentation in CT images. An analytical criterion allowing for an automatic tuning of the method is developed. The experiments are performed using a data set of 54 images from one patient, demonstrating the validity of the proposed criterion. The results are evaluated using the Jaccard index and a measure of border error with tolerance which addresses, application-dependant, acceptable error. The outcome shows that the technique has a great potential to become a method of choice for segmentation of CT images with an overall average boundary precison, for six representative tissues, equal to 0.937.

Automatic Tuning of MST Segmentation of Mammograms for Registration and Mass Detection Algorithms

A technique utilizing an entropy measure is developed for automatically tuning the segmentation of screening mammograms by minimum spanning trees (MST). The lack of such technique has been a major obstacle in previous work to segment mammograms for registration and applying mass detection algorithms. The proposed method is tested on two sets of mammograms: a set of 55 mammograms chosen from a publicly available Mini-MIAS database, and a set of 37 mammograms selected from a local database. The method performance is evaluated in conjunction with three different preprocessing filters: gaussian, anisotropic and neutrosophic. Results show that the automatic tuning has the potential to produce state-of-the art segmentation of mass-like objects in mammograms. The neutrosophic filtering provided the best performance.

Low-error, High-speed Approximation of the Sigmoid Function for Large FPGA Implementations

There has been much study of ASIC neurocomputers but, in comparison, relatively little for FPGA neurocomputers. Nevertheless, with current (and future) dense, high-speed FPGAs, the latter are now viable and will be more successful than the former. In this paper, we discuss a technique for low-error, high-speed implementations of the sigmoid function in such FPGAs. This function is commonly used as an activation function in artificial neural networks, but it also has applications in many other areas. Our results compare very favourably with others that have been reported in the published literature.

Automatic tuning of a graph-based image segmentation method for digital mammography applications

Mammogram segmentation tasks underpin a wide range of registration, temporal analysis and detection algorithms. Unfortunately, finding an accurate, robust and efficient segmentation still remains a challenging problem in mammography. A recent segmentation technique, based on minimum spanning trees (MST segmentation), is known to be robust to typical mammogram distortions and computationally efficient. This method captures both local and global image information but the balance requires choosing a parameter. So far no automatic procedure to estimate this parameter has been proposed and the value was determined experimentally. In this paper a segmentation evaluation criterion, based on a measure of image entropy, is used to automatically optimize the granularity of an MST-based segmentation. The method is tested on a set of 82 random images taken from a commonly used mammogram database. The results show a dramatic improvement in the accuracy of a MST segmentation tuned up using the entropy-based criterion.

Temporal Analysis of Mammograms Based on Graph Matching

A method is proposed for detecting masses in screening mammograms by analyzing changes between current and previous mammograms. The method uses graph matching in order to circumvent the problem of registering images of the same breast taken up to three years apart. Ninety five temporal pairs of images were separated into a training set (51 pairs) and a testing set (44 pairs). A small increase in performance, as measured by the area under the ROC curve, was found for the testing set when detection rates with graph matching were compared to detection rates without graph matching.

Robustness of Two Methods for Segmenting Salient Features in Screening Mammograms

The performance of two image segmentation methods are compared according to robustness of the segmentation to image distortion. This criterion is crucial for temporal analysis of screening mammograms where natural changes in the breast plus inherent deformation of soft tissue during image acquisition result in severe image registration problems. A method based on minimum spanning trees (MST) is found to be more robust to the distortions studied than a method based on adaptive pyramids (AP). Although segmentation leads to great differences in segmentation in distorted images for many components of low saliency, salient components (those of primary interest) are found to be segmented consistently regardless of distortion.