Does intensity windowing improve the detection of simulated calcifications in dense mammograms?

Abstract

This study attempts to determine whether intensity windowing (IW) improves detection of simulated calcifications in dense mammograms. Clusters of five simulated calcifications were embedded in dense mammograms digitized at 50-1~m pixels, 12 bits deep. Film images with no windowing applied were compared with film images with nine different window widths and levels applied. A simulated cluster was embedded in a realistic background of dense breast tissue, with the position of the cluster varied. The key variables involved in each trial included the position of the cluster, contrast level of the cluster, and the IW settings applied to the image. Combining the ten IW conditions, four contrast levels and four quadrant positions gave 160 combinations. The trials were constructed by pairing 160 combinations of key variables with 160 backgrounds. The entire experiment consisted of 800 trials. Twenty student observers were asked to detect the quadrant of the image in which the mass was Iocated. There was a statistically significant improvement in detection performance for clusters of calcifications when the window width was set at 1024 with a level of 3328, and when the window width was set at 1024 with a level cf 3456. The selected IW settings should be tested in the clinic with digital mammograms to determine whether calcification detection performance can be improved. Copyright 9 1997 by W.B. Saunders CompanyKEY WORDS: mammography, image processing, intensity windowing, observer studies, calcifications, computers, radiology.M AMMOGRAPHY, especially in women with dense breasts, is not perfectly sensitive to all cancers. Approximately 10% to 15% of palpable malignancies are not visible mammographically. ~ Tbere is some reason to believe that digital mammography rnight allow for greater contrast and improved detection of small and early tumors over standard film screen technology, especially ir image processing is used to improve image contrast. 2,3There are many potentially useful image processing algorithms, and each algorithm has a number of parameters that can be systematically varied to improve or worsen lesion detectability. Radiologists cannot and should not evaluate these algorithms in the clinic with real patients. Such a task would be overwhelming and potentially could cause much unnecessary patient anxiety. Ideally, a test set of image phantorns with simulated lesions in known locations should be used to test each potentially useful algorithm and its attendant parameters in the laboratory setting before any patient's images ate interpreted using these algorithms. We have developed such a laboratory method for evaluation of image processing algorithms. 4 In previous work, we have shown that detection performance with the application of contrast limited adaptive equalization (CLAHE) to digitized mammograms is parallel for radiologists and student observers. 4 Using the same experimental paradigm, we report here on whether intensity windowing (IW) can improve the detection of calcifications in dense mammograms in...

Evaluation of clinical image processing algorithms used in digital mammography

“…The relatively large number of cases ͑200͒ used in our study and the fact that 33% of our cases had multiple lesions also contributed to higher statistical power than earlier studies. Only two 10,16 of the seven prior studies were objective ROC studies and five [11][12][13][14][15] used very few cases ͑Յ40͒. None of the two ROC studies used optimal ROC analysis ͑e.g., DBM͒ that takes into account correlations due to the reader and case matching in the data to maximize statistical power.…”

“…10-16 and references therein͒. However, these studies are not directly applicable to modern digital mammography since one or more of the following conditions were true: the images were viewed in hard-copy mode, 10,11,[13][14][15][16] which is not the current standard as it does not take into account the flexibility of user-adjustable window and level controls in soft-copy display; the studies utilized subjective preference methods ͓i.e., non-ROC/FROC ͑free-response receiver operating characteristic͔͒ that cannot control for observer reporting threshold variability; 11,12 only effects of window-level ͑W/L͒ adjustments were studied using student observers, 13,14 whereas in practice W/L is the final step of a series of complex image processing steps performed on the raw data and radiologists are the final observers; they used in-house developed algorithms [11][12][13][14][15] rather than manufacturer-delivered algorithms, such as are used in the clinics; and none of the prior studies used modern ROC/ FROC evaluation methodology. 17,18 The aim of the current study was to determine if manufacturer-developed image processing algorithms, which are actually used under clinical conditions in a fully digital mammography reading environment at our institution, have an effect on the interpretation of digital mammograms.…”

Radiologists’ Preferences for Digital Mammographic Display

“…Undoubtedly, habit and experience influenced the preference of radiologists for screen-film images over processed digital images in many cases. A prior preference study (4), in which the investigators attempted to exactly match the appearance of the screen-film mammograms through MIW, showed that radiologists preferred digital mammography to screen-film imaging. Of course, such matching might not allow the full benefits of digital mammography to be realized.…”

Interobserver agreement and performance score comparison in quality control using a breast phantom: screen-film mammography vs computed radiography

“…As for image quality, the accuracy of CR reading is affected by the image-processing parameters and windowing [7,8,9]. Kallergi et al [10] reported that interobserver variation in the digitized mammogram reading (105-µm pitch, 12-bit gray scale) was greater than in the film reading.…”