Gray-scale and geometric registration of full-field digital and film-screen mammograms
Introduction
In radiology, the comparison of images obtained in subsequent examinations of a patient often is an important part of the diagnostic procedure. These comparisons are made to detect interval changes indicating lesion growth, to monitor progression of a disease, or to estimate the effect of treatment. Sometimes, temporal images are subtracted to enhance areas where differences occur. In conventional radiology, the review of temporal image pairs may be seriously hampered by differences in acquisition. To some extent, geometric registration algorithms, the development of which received a lot of attention in recent years (Sallam and Bowyer, 1999, Wirth et al., 2002, Marias et al., 2005, Pluim, 2003) can deal with positioning changes. On the other hand, changes of exposure and detector systems may also reduce the effectiveness of temporal comparisons, because of the nonlinear gray-scale changes they may induce.
One area where comparison with prior imaging plays a major role is breast cancer screening. Studies have shown that the use of prior mammograms in screening effectively reduces the number of false positive referrals (Thurfjell et al., 2000, Burnside et al., 2002). This is because priors allow radiologists to distinguish growing lesions from normal dense structures in the breast that somehow look suspicious. Differences between acquisition/display systems and between exposures may cause subsequent mammograms to appear dramatically different, which is annoying for the radiologists and may reduce their performance. These differences cannot be corrected during display as long as conventional alternators are used for reading. With the introduction of digital mammography and dedicated mammographic workstations, however, the problem of display optimization can be addressed properly.
During the transition to digital mammography, hospitals and mammography screening centers are confronted with mammographic cases that both contain old FS mammograms and newly taken FFD mammograms. Especially, in countrywide screening programs, this issue cannot be trivialized because of the quantities of mammograms that are produced and because of practical issues of workflow and reading design. Depending on the screening interval, the transition period is 1–3 years when one prior mammogram is used for comparison. Sometimes, radiologists also like to see mammograms one screening round before that, which doubles the transition period.
Combined usage of an alternator and a dedicated mammographic workstation is one possibility to read hybrid cases, but this has practical drawbacks that probably stand in the way of good image comparison. The main drawback is that the alternator produces high luminance, while soft-copy reading demands dimmed lights. Another possibility is reading both the prior and current mammograms on an alternator, after hard-copies are made of the FFD mammograms. This has the drawback that the benefits of FFD mammography (archiving, transmission, remote reading, computer-aided detection, and image enhancement) are not fully exploited for a number of years more. Mostly, FS mammograms will be digitized so that both images can be read and compared by soft-copy reading. This choice gives the benefits of soft-copy reading, while the quality of reading does not need to be impaired (see also Roelofs et al., 2006, Pisano et al., 2005). Optimizing the display of hybrid mammogram pairs to facilitate comparison is the focus of this paper. This paper presents a model-based method that allows matching digitized FS mammograms to unprocessed FFD mammograms both geometrically and in gray-scale. Although the primary objective is image display, it is remarked that the technique will also be important in the development of computer-aided detection methods that make use of temporal comparison.
Histogram matching is the generic term of all methods that match images with the same or similar content by adjusting lookup tables of pixel values. Hence, histogram matching is typically a one-dimensional process. In general, histogram-matching methods can be divided into two classes: (1) non-parametric methods that try to find pairs of pixel values that match as close as possible and (2) parametric methods that, as the word says, require the estimation of parameters. The latter are often based on transformations with simple polynomial functions. The method, presented here, is a member of the parametric matching methods. The novelty of this work is that a particular model-theoretic form for the gray-scale transform is derived based on an understanding of mammographic image formation: it models compression of the breast; exposure time; radiation intensity of the incident beam; for analog images it models the sensitometric curve and the process of digitization; and, finally, for unprocessed digital images it models the relationship between exposure and pixel value. The chain of events in acquisition defines the objective transform between the mammograms. The motivation to incorporate aspects of image formation is to prevent the introduction of spurious properties of the transformed image that might put radiologists on the wrong track; though important, minimizing residual errors in pixel values is subordinate to this. Another difference with other parametric methods for histogram matching is the computational procedure to determine/optimize the parameters of the transform, namely the registration algorithm. Conventionally, the registration algorithm optimizes a figure-of-merit function that quantifies the difference between two distributions of pixel values, while we use a figure-of-merit function that quantifies differences at the level of pixel pairs. Therefore, we rather speak of gray-scale registration than of histogram matching to make the distinction.
In this paper, the geometric registration to find pixel pairs is simultaneously performed with the gray-scale registration. The transformation consists of rotation, translation, and isotropic scaling of one of the images. The transformation can be viewed as merely a by-product of gray-scale registration, but it can also be given a more prominent role when the geometric transformation is enabled for displayed images. Especially, while toggling between prior and current mammograms, which is sometimes done to assess lesion growth; radiologists reported that the images were easier to read when they were geometrically registered (see also van Engeland et al., 2003).
Snoeren and Karssemeijer (2003) first introduced the gray-scale transform between FS and FFD mammograms. Since then the registration algorithm has undergone major improvements with respect to the computation time and, more importantly, the capture range of transformation parameters. This is accomplished by simultaneous geometric and gray-scale registration, a different optimization technique that utilizes gradient information of the figure-of-merit function, and a coarse-to-fine strategy.
Section snippets
Theory
The gray-scale and geometric transform, which will be derived in the first part of this section, is tailored to the transformation of a FS mammogram to an unprocessed FFD mammogram. The most relevant aspects of acquisition are modeled for both modalities. This encompasses (1) breast positioning; (2) breast compression; (3) exposure time; (4) incident radiation intensity; and (5a) film properties and digitization for FS mammograms, or (5b) detector response for FFD mammograms. The resulting
Results
We presented 40 complete cases with current and prior mammograms of both breasts to radiologists. All FFD mammograms were from the General Electric Senographe 2000D, a mammography system with a linear detector response. Comparing the cases before and after registration, the radiologists reported that the gray-scale registered mammogram pairs were excellent. We realize that this assertion depends on how digitized FS mammograms and FFD mammograms are normally displayed. Namely, there is no
Discussion
One aspect of the registration method is the parametric gray-scale transform between the two images, which is specifically designed for FS–FFD mammogram pairs. Among transforms for the other modality pairs, i.e., FS–FS and FFD–FFD, this transform was already introduced in 2003 by Snoeren and Karssemeijer (2003). In the current paper, we focused on the other aspect of the registration method, the registration algorithm. It comprises a geometric registration that is carried out simultaneously
Conclusion
A method of parametric geometric and gray-scale registration of mammogram pairs is presented, which is specifically designed for combinations of an unprocessed full-field digital and a film-screen mammogram of the same breast. The novelty of this work is that a particular model-theoretic form for the gray-scale transform is derived based on an understanding of mammographic image formation.
This paper focused on mammography, but film-screen technique also has passed its best to other fields of
References (15)
Fast filter transforms for image processing
Computer Graphics and Image Processing
(1981)- et al.
Registration and difference analysis of corresponding mammogram images
Medical Image Analysis
(1999) - et al.
Differential value of comparison with previous examinations in diagnostic versus screening mammography
American Journal of Roentgenology
(2002) - et al.
The Laplacian pyramid as a compact image code
IEEE Transactions on Communications
(1983) - et al.
A registration framework for the comparison of mammogram sequences
IEEE Transactions on Medical Imaging
(2005) - Part 14: Grayscale Standard Display Function. Digital Imaging and Communications in Medicine (DICOM). National...
- et al.
Image processing algorithms for digital mammography: a pictorial essay
Radiographics
(2000)
Cited by (11)
Computer-aided detection of masses in full-field digital mammography using screen-film mammograms for training
2008, Physics in Medicine and BiologyLatent feature representation with 3-D multi-view deep convolutional neural network for bilateral analysis in digital breast tomosynthesis
2016, ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - ProceedingsFeature extraction from bilateral dissimilarity in digital breast tomosynthesis reconstructed volume
2015, Proceedings - International Conference on Image Processing, ICIPDetection and diagnosis of breast masses in mammography
2015, Computer-Aided Detection and Diagnosis in Medical ImagingBreast image registration and deformation modeling
2012, Critical Reviews in Biomedical EngineeringA statistical approach for breast density segmentation
2010, Journal of Digital Imaging