State of the Art in Digital Mammographic Image Analysis

The following sections are included:

• EDITOR'S INTRODUCTION

• CONTENTS

Mammographic screening programmes generate large numbers of highly variable, complex images, most of which are unequivocally normal. When present, abnormalities may be small or subtle. Two processes critical to the success of screening programmes are the perception of potential abnormalities and the subsequent analysis of each detected lesion to determine its clinical significance. The consequences of errors are costly, and in many screening centres, films are read by two radiologists in an attempt to reduce errors. The prime objective of our research is to improve the accuracy of the detection and analysis of breast lesions by providing radiologists with computer-aided digital image analysis tools. In this paper we focus on the detection and analysis of mammographic microcalcifications.

We describe a philosophy of research aimed at generating useful computer-based aids for radiologists. Firstly, it is necessary to accurately identify specific tasks which are difficult for the human observer. Having correctly identified a problem, appropriate computer vision methods must be developed and their performance evaluated. It is then important to determine effective ways of using such methods to aid radiologists, and it is essential to prove that the effect on radiologists' performance is entirely beneficial.

We present results of experiments to determine factors affecting radiologists' perception of microcalcifications, and to investigate the effects of attention-cueing on detection performance. Our results show that radiologists' performance can be significantly improved with the use of prompts generated from automatically-detected microcalcification clusters.

We describe a new method for the delineation of mammographic abnormalities based on the analysis of multiple high quality X-ray projections of excised lesions. Biopsy specimens are secured inside a rigid tetrahedron, the edges of which provide a reference frame to which the locations of features can be related. A three-dimensional representation of an abnormality can be formed and rotated to resemble its appearance in the original mammogram.

The following sections are included:

• Introduction

• Mammographic Features
  • Opacities
  • Calcifications

• Optimal Sampling of Mammograms

• Senobase

• Transform Image Coding

• Adaptive Coding Method for Mammograms
  • Thresholding
  • Coding the Location of the DCT
  • Coding the Amplitude of the DCT
  • Properties of the Whole Coding Scheme

• Results and Discussion

• Results - Discussions

• References

Previous efforts in improvement of appearance of digitized mammograms have concentrated on using image enhancement techniques. In this work we propose the application of image restoration.

We considered a non-stationary image model and signal-dependent noise of photonic and film-grain origins. Both the camera blur and the MTF of the screen-film combination were considered. The camera noise was minimized through averaging and background subtraction. The signal-dependent nature of the radiographic noise was modeled by a linear shift-invariant system and the relative strengths of various noise sources were compared.

We developed and implemented two locally adaptive image smoothing filters to improve the signal to noise ratio of digitized mammogram images. To minimize the effects of the system blur a deconvolution filter was then applied in conjunction with these smoothing filters resulting in better visualization of image details.

The deconvolution filter was based on the Minimum Mean Squared Error (MMSE) criteria, while the smoothing filters utilize the Baysian and the Wiener criteria. Of the two smoothing filters the Baysian estimator was found to outperform the adaptive Wiener filter. The filters were implemented in a real time processing environment using our mammographic image acquisition and analysis system.

We are investigating techniques for automatically sorting mammograms according to whether the breast tissue is fatty or dense. The hypothesis is that areas of dense tissue are a major factor in making certain mammograms harder for both radiologists and computers to interpret. Being able to identify dense mammograms automatically would enable expert radiologists to use their time and skills more efficiently, since only the difficult mammograms would be examined by the most experienced readers. Concentrating on the easier, fatty mammograms might make the computer-aided detection of abnormalities more tractable.

A number of local statistical and texture measures have been applied to manually selected patches from digitised mammograms. One of the measures (local skewness in tiles) gives a good separation between fatty and dense patches. A method of automatic patch placement has been devised and a fully automatic procedure for sorting mammograms is therefore possible.

The following sections are included:

• Introduction

• Image Digitization and Breast Segmentation

• Detection of Clustered Microcalcifications
  • Method
  • Performance
  • Application of Artificial Neural Networks to the Detection of Clusters

• Detection of Breast Masses
  • Method
  • Performance

• Classification of Breast Lesions

• “Intelligent” Mammography Workstation

• Discussions
  • Case Selection
  • Scoring Method
  • Performance Evaluation Method

• Summary

• Acknowledgments

• References

This paper describes the design, implementation, and testing of an adaptive digital image segmentation method that detects cancerous changes in mammograms and can potentially aid medical experts in establishing the diagnosis. The essence of the method is hierarchical region growing that uses pyramidal multiresolution image representation. The relationships between pixels at different resolution levels are established using a fuzzy membership function, thus enabling detection of very small and/or low contrast objects in a highly textured background. The selection of the parameters of the fuzzy membership function allows for fine-tuning the method to specific segmentation objectives. This paper discusses two versions of the method: the first is aimed at the detection of microcalcifications and the second at the detection of benign and malignant nodules. The two versions are fully automated and differ in the procedure applied to automatically select the appropriate parameters of the fuzzy membership function. Both versions were evaluated in two ways: (i) using synthetically generated objects superimposed on normal mammograms and (ii) using mammogram images for which the corresponding truth images were generated by human experts. The objective of the first evaluation was to precisely determine the method's capabilities and its sensitivity to object size, shape, and contrast. The objective of the second evaluation was to establish the method's usefulness in helping medical experts to establish the diagnosis.

A framework for computer-aided analysis of mammograms is described. General computer vision algorithms are combined with application specific procedures in a hierarchical fashion. The system is under development and is currently limited to detection of a few types of suspicious areas.

The image features are extracted by using feature extraction methods where wavelet techniques are utilized. A low-pass pyramid representation of the image is convolved with a number of quadrature filters. The filter outputs are combined according to simple local Fourier domain models into parameters describing the local neighbourhood with respect to the model. This produces estimates for each pixel describing local size, orientation, Fourier phase, and shape with confidence measures associated to each parameter.

Tentative object descriptions are then extracted from the pixel-based features by application-specific procedures with knowledge of relevant structures in mammograms. The orientation, relative brightness and shape of the object are obtained by selection of the pixel feature estimates which best describe the object.

The list of object descriptions is examined by procedures, where each procedure corresponds to a specific type of suspicious area, e.g. clusters of microcalcifications.

A statistical method is described for detection of microcalcifications in digital mammograms. It is shown that the detection performance depends strongly on a preprocessing step, in which the images are rescaled to equalize image noise. A robust algorithm is proposed for rescaling, which can be used to determine a proper scale conversion from a phantom recording. The same algorithm, however, can also be applied to the image to be processed itself. Such an adaptive approach, in which noise characteristics are estimated from the image at hand, appeared to be the basis for far better results than could be obtained by using a fixed scale conversion. The method used for detection is based on Bayesian techniques. A random field model is used to model spatial relations between the labels in an iterative segmentation process. Results of an experimental study using a set of 65 mammographic images digitized at 2048 × 2048 are presented.

Mammography associated with clinical breast examination and breast self-examination is the only effective and viable method for mass breast screening. Most of the minimal breast cancers are detected by the presence of microcalcifications. It is however difficult to distinguish between benign and malignant microcalcifications associated with breast cancer. Most of the techniques used in the computerized analysis of mammographic microcalcifications segment the digitized gray-level image into regions representing microcalcifications. Since mammographic images usually suffer from poorly defined microcalcification features, the extraction of microcalcification features based on segmentation process is not reliable and accurate. We present a second-order gray-level histogram based feature extraction approach which does not require the segmentation of microcalcifications into binary regions to extract features to be used in classification. The image structure features, computed from the second-order gray-level histogram statistics, are used for classification of microcalcifications. Several image structure features were computed for 100 cases of “difficult to diagnose” microcalcification cases with known biopsy results. These features were analyzed in a correlation study which provided a set of five best image structure features. A feedforward backpropagation neural network was used to classify mammographic microcalcifications using the image structure features. Four networks were trained for different combinations of training and test cases, and number of nodes in hidden layers. False Positive (FP) and True Positive (TP) rates for microcalcification classification were computed to compare the performance of the trained networks. The results of the neural network based classification were compared with those obtained using muitivariate Baye's classifiers, and the k-nearest neighbor classifier. The neural network yielded good results for classification of “difficult-to-diagnose” microcalcifications into benign and malignant categories using selected image structure features.

We propose a detection and classification system for the analysis of mammographic calcifications. First, a new multi-tolerance region growing method is proposed for the detection of potential calcification regions and extraction of their contours. The method employs a distance metric computed on feature sets including measures of shape, centre of gravity, and size obtained for various growth tolerance values in order to determine the most suitable parameters. Then, shape features from moments, Fourier descriptors, and compactness are computed based upon the contours of the regions. Finally, a two-layer perceptron is utilized for the purpose of classification of calcifications with the shape features. A new leave-one-out algorithm-based parameter determination procedure is included in the neural network training step. In our preliminary study, detection rates were 81% and 85 ± 3%, and correct classification rates were 94% and 87% with a test set of 58 benign calcifications and 241 ± 10 malignant calcifications, respectively. The proposed system should provide considerable help to radiologists in the diagnosis of breast cancer.

Computer-assisted detection of microcalcifications in mammographic images will likely require a multi-stage algorithm that includes segmentation of possible microcalcifications, pattern recognition techniques to classify the segmented objects, a method to determine if a cluster of calcifications exists, and possibly a method to determine the probability of malignancy. This paper focuses on the first three of these stages, and especially on the classification of segmented local bright spots as either calcification or noncalcification. Seven classifiers (linear and quadratic classifiers, binary decision trees, a standard backpropagation network, 2 dynamic neural networks, and a K-nearest neighbor) are compared. In addition, a post-processing step is performed on objects identified as calcifications by the classifiers to determine if any clusters of microcalcifications exist. A database of digitized film mammograms is used for training and testing. Detection accuracy of individual and clustered microcalcifications is compared across the seven methods using area under the ROC curve as a figure of merit.

The following sections are included:

• Introduction

• Data Acquisition

• Automatic Classification
  • Feature extraction
  • Classification
  • Classifier optimisation

• Results

• Discussion

• References

• Appendix: definition of features

Breast asymmetry is an important radiological sign of cancer, this paper describes the first approach aiming to detect all types of asymmetry; previous asymmetry-based research has been focussed on the detection of mass lesions. The conventional approach is to search for brightness or texture differences between corresponding locations on left and right breast images. Due to the difficulty in accurately identifying corresponding locations, asymmetry cues generated in this way are insufficiently specific to be used as prompts for small and subtle abnormalities in a computer-aided diagnosis system. We have undertaken studies to discover more about the visual cues utilized by radiologists. As a result, we propose a new automatic method for detecting asymmetry based on the comparison of corresponding anatomical structures, identified by an automatic segmentation of breast tissue types. We describe methods for comparing the shape and brightness distribution of these regions, and we present promising results obtained by combining evidence for asymmetry.

We have previously reported on a method for the automatic detection of stellate lesions in digitized mammograms, and on our tests of that method on image data with known diagnoses. This earlier investigation was based on a limited set of 10 test images, each with a stellate lesion. As our approach is one of supervised training, half of the data was used as a training set, and so the performance results were necessarily coarse.

Accordingly there is value in testing these algorithms on a larger data set that will not only provide more lesions but also truly undiseased tissue. A new mammogram data set addresses both of these concerns, as it contains examples of twelve stellate lesions, as well as fifty examples of entirely normal mammograms. Further, as this data set has been made widely available to all interested researchers, performance results for specific algorithms on this data set are of particular value, as they can be directly compared to the performance of other algorithms similarly applied.

Thus the main contribution of the current paper is to exhaustively evaluate the performance of this stellate lesion detection algorithm on the new mammogram data set. A secondary aim is to present a revision of the spatial integration step which generates the final report of a lesion's existence, one that facilitates the extraction of ROC performance statistics.

The following sections are included:

• Introduction

• Greyscale Image Processing: Summary of Previous Work

• Computer-Aided Diagnosis: Summary of Previous Work

• Greyscale Image Processing and Computer-Aided Diagnosis: Future Work

• Conclusion

• References