Catálogo de publicaciones - libros

The volume of dense breast tissue can be calculated from an x-ray mammogram by imaging a calibrated step-wedge alongside the breast and determining the compressed breast thickness. Previously published work used a step-wedge made of PTFE with a maximum height of 35mm, length 175mm and width 15mm. Although fulfilling all theoretical requirements, it can be difficult to find space on the film for a large step-wedge when examining bigger breasts. Furthermore, the step-wedge is lead-lined, making it heavy and difficult to attach to the bucky. A more compact aluminium step-wedge has been designed to overcome these limitations, and experiments have been carried out on a prototype to evaluate its performance. Initial results show that the maximum and minimum heights of the prototype step-wedge are inadequate to sufficiently cover the range of optical densities within a breast image at the higher and lower exposures required for 6cm and 2cm Perspex (>200mAs and <40mAs respectively). However, the step increment appears to be satisfactory. Analysis of the mean pixel value and standard deviation within Regions of Interest of varying size and position indicates an optimum step length of 3mm. A new step-wedge has been constructed with an improved specification informed by the evaluation of the prototype.

In medical image analysis a ground truth to compare results against is of vital importance. This ground truth is often obtained from human experts. The aim of this paper is to discuss the problem related to the use of markings made by an expert panel. As a partial solution, we propose a method to relate markings to each other in order to establish levels of agreement. By using this method we can assess the performance of, for instance, segmentation algorithms.

Methods for improving the accuracy of a technique for estimating volumetric breast density are described. A breast tissue-equivalent phantom encompassing a range of thicknesses and compositions of tissue is used to evaluate the sources of error in the technique. The image acquisition parameters that can affect the accuracy of calibration are considered, and sensitivity to these factors is evaluated. The robustness of the technique was tested by obtaining calibration images on 24 mammography machines, at 18 different sites, over a period of 3 years. The ability to use a single calibration on all machines of a given model type was assessed by comparing effective linear attenuation coefficients of fat and fibroglandular tissues, derived from the calibration phantom images obtained from various machines.

The effect on the measurement of volumetric breast density of variations in physical and chemical properties of adipose and fibroglandular tissue reported in a number of studies is investigated using the authors’ model of mammographic image formation. This model is developed specifically for the measurement of breast density. The effect of varying stromal composition, a popular histopathological explanation of mammographic density, is also discussed. Given the uncertainties in tissue attenuation highlighted by this study, as well as noise, and acquisition model error, the validity of this measurement is discussed, together with alternative measurement scales. Several issues are considered, including the effect of beam quality on normalisation accuracy, and the measurement failure which can occur when clinical data falls outside the limited range defined by 100% adipose to 100% fibroglandular tissue.

This paper describes initial steps towards the development of a Computer Aided Detection (CAD) system based on breast density pattern classes. We present evidence that the sensitivity and specificity of such a system will improve if it is developed for, and applied to, specific breast density classes.

Breast density segmentation and classification methods are combined to enable the automatic and quantitative comparison of temporal mammograms of women using Hormone Replacement Therapy (HRT). The results are based on registration and density quantification, so that potentially the clinician may be informed about substantial localised breast density changes. The measures use texture based density segmentation as well as a normalized representation of mammograms.

We have developed a Monte Carlo model to examine the cancer detection rate in screening mammography. We simulated the situation where screening was implemented for 9 years and then CADe was implemented for an additional 9 years. We investigated the effectiveness of two different methods for measuring changes in cancer detection rate. The first method was a sequential method in which the radiologist first reads without CADe and then immediately reads with CADe. The second method is temporal comparison where the cancer detection rates for two periods of time are compared: one without the use of CADe and one when CADe is in use. The model predictions have important implications for clinical studies of CADe. The temporal method is unlikely to measure a real affect, because the effect is small. A sequential method can measure an increase in the number of cancers detected because of CADe, but it cannot measure an overall increase in the cancer detection rate of the screening program.

Systems for computer aided detection of masses may be used more effectively when they are used for interpretation of suspect abnormalities, instead of solely using them as a prompting aid to avoid oversights. To use CAD algorithms for detection of masses as a decision aid it may be helpful to display suspiciousness of regions computed by CAD. In this paper the quality of probabilities computed for masses by a commercial CAD system is studied in two ways: 1) by comparing standalone performance of the system to that of experienced screening radiologists, and 2) by determining results of independent double reading with CAD. The study involves results of 15 readers who each read 500 mammograms, and two releases of the CAD algorithm. Independent double reading results are obtained by combining probabilities of the CAD system with the reader assessment for each localized finding reported by the reader, and by computing the fraction of cancers localized correctly as a function of false positive referrals. It was found that standalone performance of CAD is less than that of any reader in the study. Nevertheless, it was found that performance improves significantly with independent CAD reading, and that use of an improved CAD algorithm lead to significantly better results of the combined reader with CAD.

The UK National Health Service Breast Screening Programme (NHSBSP) provides free mammographic screening for all women between the ages of 50 and 69. This paper examines in detail the way in which the programme is implemented in one of the busiest breast screening centres, discussing the implications of current practice for the introduction of computer aided detection systems. The paper also investigates the different types of abnormality that arise in older and younger women within the screening age group, and discusses how this is likely to affect prompting systems.

We develop a novel CAD detection system that can help a radiologist to detect masses in mammograms. The proposed algorithm concurrently detects the breast boundary and the pectoral muscle. Then, a clustering and morphology based segmentation algorithm is applied to the enhanced mammography image to separate the mass from the normal breast tissues. This technique outlines the shape of candidate masses in mammograms. To maximize detection specificity, we develop a two-stage hybrid classification network. First, an unsupervised classifier is used to classify suspicious opacities as suspect or not. Then, a few supervised interpretation rules are applied to further reduce the number of false detections. Using a private mammography database and the publicly available USF/DDSM database, experimental results demonstrate that a sensitivity of 94% (resp. 80%) can be achieved at a specificity level of 1.02 (resp. 0.69) false positives per image. Even in dense mammograms, the CAD algorithm can still correctly detect subtle masses.