A Novel Software-Based Technique for Quantifying Placental Calcifications and Infarctions from Ultrasound
John T Ryan*a, Fionnuala McAuliffeab, Mary Higginsab, Marie Stantona, Patrick Brennana a School of Medicine and Medical Science, University College Dublin, Dublin 4, Ireland; b National Maternity Hospital, Holles St., Dublin 2, Ireland.
ABSTRACT In obstetrics, antenatal ultrasound assessment of placental morphology comprises an important part of the estimation of fetal health. Ultrasound analysis of the placenta may reveal abnormalities such as placental calcification and infarcts. Current methods of quantification of these abnormalities are subjective and involve a grading system of Grannum stages I-III. The aim of this project is to develop a software tool that quantifies semi-automatically placental ultrasound images and facilitates the assessment of placental morphology. We have developed a 2D ultrasound imaging software tool that allows the obstetrician or sonographer to define the placental region of interest. A secondary reference map is created for use in our quantification algorithm. Using a slider technique the user can easily define an upper threshold based on high intensity for calcification classification and a lower threshold to define infarction regions based on low intensity within the defined region of interest. The percentage of the placental area that is calcified and also the percentage of infarction is calculated and this is the basis of our new metric. Ultrasound images of abnormal and normal placentas have been acquired to aid our software development. A full clinical prospective evaluation is currently being performed and we are currently applying this technology to the three-dimensional ultrasound domain. We have developed a novel softwarebased technique for calculating the extent of placental calcification and infarction, providing a new metric in this field. Our new metric may provide a more accurate measurement of placental calcification and infarction than current techniques. Keywords: Placental Calcification, Ultrasound, Quantification, Segmentation
1. INTRODUCTION Radiological imaging, a fundamental component of health care systems and laboratory research, is undergoing unprecedented change where imaging modalities and analytical tools are being developed to display biological tissues and processes in novel quantitative ways. These developments are dependent on multidisciplinary imaging research collaborations, involving clinicians, biologists, computer specialists, biomedical engineers and visual scientists to support, validate, interpret and optimize current and new imaging methods. Health-related imaging research is a stated priority of UCD1, Irish Government2,3 and European Commission4.
*
[email protected]; phone +353 1 7166539; fax +353 1 7166547; http://www.ucd.ie/diagnosticimaging/html/johnryan Medical Imaging 2008: Ultrasonic Imaging and Signal Processing, edited by Stephen A. McAleavey, Jan D'hooge, Proc. of SPIE Vol. 6920, 69200P, (2008) · 1605-7422/08/$18 · doi: 10.1117/12.769464
Proc. of SPIE Vol. 6920 69200P-1 2008 SPIE Digital Library -- Subscriber Archive Copy
Computer-Aided Detection (CAD) and segmentation algorithms combined with computer visualisation techniques have revolutionized diagnosis and treatment in many areas of medicine. These algorithms and techniques allow the clinician to assess and evaluate structure as well as quantitatively defining abnormalities. Ultrasound has always posed problems for automatic segmentation and CAD due to the noise that is prevalent in this imaging modality. As technology progresses with sophisticated filters and noise removal algorithms, the impact of noise on ultrasound images is becoming less pronounced. Furthermore, the operator-dependant nature of ultrasound imaging coupled with the increasing popularity in expert-seeded algorithms strengthen our existing and future software approaches. We have developed a software tool for the analysis of placental calcification and infarction. In the primary instance, this novel tool has been developed for placental calcification which are visualized with ultrasound prenatal examinations. Calcification of the placenta is associated with foetal distress in labour5, poor perinatal outcome6, maternal smoking7, preeclampsia8; all of which result in increased risk of fetal and neonatal morbidity and mortality6,9,11. Grannum classification is the currently employed method for describing this important abnormality, which is based on a subjective observation of the placenta and depending on the presence and location of the calcifications a score of I, II or III is given with the larger number describing a greater presence of calcification. The lack of objectivity, precision and reproducibility of the current method at least in part contributes to the lack of progress identifying a high risk foetus in a low risk population12. In the western world, almost all women are offered at least one ultrasound examination during pregnancy. The Euronatal audit study10 demonstrated that accurate detection of growth retardation in-utero, a marker of which is placental calcification, can result in better management of pregnancies with the implementation of interventions that reduce perinatal death9. This is clearly an important clinical area, on which to test the applicability and usability of any quantification tool. The tool will be further developed to demonstrate automatic quantification of other indicators of foetal development such as foetal abdominal fat content.
2. SYSTEM OVERVIEW We have developed a 2D ultrasound imaging software tool that allows the obstetrician and sonographer to define the placental region of interest. Using a slider the user can easily define an upper threshold based on high intensity for calcification classification and a lower threshold to define placental infarction areas based on low intensity within the defined region of interest. The percentage of the placental area that is calcified and also the percentage of infarction is calculated and this is the basis of our new metric. Ultrasound images of abnormal and normal placentas have been acquired in collaboration with, the National Maternity Hospital, Dublin. Preliminary results and feedback suggest that this tool may provide a greater range in the scoring of placental calcification than the traditional methods. Further development is needed to accurately quantify percentage of placental infarction within the placenta. We are currently validating the preliminary software approach in a pilot prospective clinical study. This clinical study examines the correlation between placental calcification and infarction, with perinatal outcome in postdates pregnancies. Institutional ethics approval has been obtained and with written patient consent placental images and pregnancy outcome data have been obtained. To date over 100 patients have been recruited.
Proc. of SPIE Vol. 6920 69200P-2
Selection of ROl
Flood filling from origin, stopping at red pixels
Fig. 1. Reference map creation.
Using the C++ programming language and the ITK (Insight Toolkit) and FLTK (Fast-Light Toolkit) libraries we have developed a user interface that allows the clinician to select a region of interest (ROI) to assess. Once the clinician has finished drawing this ROI, a reference map is automatically created. Figure 1 demonstrates this process. Firstly, a floodfilling algorithm is used to saturate the image (blue) apart from the red-defined ROI and it’s contents. Once this image is created, any pixels that are blue are turned black and any other pixels (i.e., within the confines of the ROI) are turned white. This now acts as a reference image. The clinician has two sliders that alter the intensity threshold for defining placental calcification and infarction within the ROI. Output metrics are in the form of pixel counts for both calcification and infarction as well as the respective percentages in reference to the total number of pixels within the ROI. Manual classification of the placental region of interest was needed due to the varying nature of the placental borders. To aid in the classification of these placental borders the Bresenham’s line drawing algorithm was used13. The line width was set to one pixel. To combat erratic or jagged hand movements of the user a smoothing algorithm was implemented. This algorithm iterates through the line vertices averaging five vertices and then averaging this result with the middle vertex to finally get a smoothed version of the original middle vertex. Due to extraneous and inherent noise associated with ultrasound images a filter was implemented. The filter works on spatial level by removing segmented objects that are defined as too small to be relevant to the quantification. A threshold, which after referencing and testing a variety of different images, was set at 10 pixels. This filter dramatically enhances the quantification in noisy images, only leaving the relevant placental calcifications. The algorithm works by traversing the segmented image until a target pixel (red) is identified. From this pixel a recursive function is called that uses a region-growing algorithm to find the extent of that particular object. When the region is fully grown, the size is checked against the defined threshold. If it is below the threshold, the pixels are turned back to the original greyscale values. If the pixels are above the threshold, these are quantified. Figures 2 to 4 demonstrate this process.
Proc. of SPIE Vol. 6920 69200P-3
Fig. 2. Unfiltered, unsegmented ultrasound image (cropped) demonstrating calcification on the surface of the placenta.
—
Fig. 3. Same image showing segmentation using thresholding technique (red)
Fig. 4. Same segmented image with white representing the objects that are included in the quantification and red representing the objects that were below the thresholded size.
Proc. of SPIE Vol. 6920 69200P-4
Fig. 5. Infarction Classification using a circular region of interest (left), blue indicates the segmented / quantified area (right). The ability to analyse and measure individual areas of placental infarction is clinically important and is not considered in the current gold-standard of the Grannum grading system. Individual placental infarction selection is enabled by implementing a circular region of interest tool. The clinician can select a central point within the infarction and grow the circle until the borders of the infarct are included (Figure 5, left). The intensity threshold slider is then used until the clinician is satisfied the infarction is full segmented (Figure 5, right). This function has value in monitoring the progress of an existing placental infarction on an individual basis rather than using the entire placental area which currently gives some false positives due to placental lakes.
3. RESULTS We have developed a software tool that uses the clinician as the key input in defining the region of interest (ROI) and the intensity thresholds. We consider this to be the best initial approach to the problem due to the insensitivity of computer aided detection (CAD) algorithms in this area and the inherent inconsistencies in ultrasound images. The user-dependant nature of ultrasound imaging reinforces our approach and preliminary clinical evaluations have proven successful. A prospective study is currently being performed to fully evaluate this tool.
Fig. 6. Screenshot of our current pilot software tool, With placental ROI, red representing calcified areas.
Proc. of SPIE Vol. 6920 69200P-5
Figure 6 demonstrates a screenshot of the user selected region of interest with the detected placental calcifications appearing as red. Figure 7 is the corresponding quantification output for the above image.
Quantification: % Calcification: 2.163751% 449 pixels out of 20751 % Infarction: 0.346971 72 pixels out of 20751 Fig. 7. Corresponding Output.
4. DISCUSSION One of the main areas of debate from this preliminary research is that of segmentation. There have been many approaches in the segmentation of ultrasound images and datasets. These have included thresholding14, region growing15– 17 , classification18, clustering19, mathematical morphology20, wavelet analysis21, genetic and fuzzy algorithms22, 23. Intensity thresholding, which is one of our current methods, can be effective in the segmentation of images/datasets containing regions with intensity values that exhibit small variance and do not overlap with the intensity values of another region. Some more sophisticated thresholding methods for segmenting ovarian cysts have been described14. This method uses a linear combination of gray level and local entropy to delineate ovarian cysts surrounded by soft tissue. However, ultrasound images commonly contain inhomogeneous and noisy/grainy regions which may cause thresholding approaches to fail24. A more complex method of segmentation called “region growing” uses an initial seed (x & y coordinates) and grows the segmented region according to a given set of rules or criteria15–17. However, ultrasound images are generally susceptible to grainy noise, tissue texture and artefacts which may lead to boundary discontinuities, small holes in the delineated regions, as well as to the delineation of incorrect regions25, 26. In addition, these approaches always need some kind of human interaction in the initiation of the seed pixels18,19. There are many other complex mathematical approaches to the segmentation of ultrasound images including mathematical morphology20, wavelet analysis21, genetic and fuzzy algorithms22, 23. Level set segmentation methods appear to be one of the strongest contenders for the accurate segmentation of ultrasound images29-31. Because placental calcification appears as multiple small nodular regions of higher intensity, the current method of simple ROI classification and thresholding combined with a filter appears to be one of the strongest solutions for this problem. To evaluate this current method, a prospective investigation will be established where 100 patients with postdates pregnancies (term plus 10 days) will undergo fetal ultrasound assessment, the placental images will be scored using the standard Grannum grading and the novel quantification method and these scores will be correlated to clinical measures of newborn well-being such as the Apgar score, cord pH, admission to the neonatal intensive care unit. Reproducibility studies will be undertaken to establish the objectivity of this technique compared with the standard subjective assessment of placental morphology. A comparison of the predictiveness of early infant morbidity with each of the two scoring techniques will be performed.
Proc. of SPIE Vol. 6920 69200P-6
5. CONCLUSIONS We have developed a software tool that quantifies the percentage of placental calcification and infarction within a userdefined ROI. A prospective study with a large patient population is currently being performed. This tool will then be further developed to demonstrate automatic quantification of other indicators of fetal development such as bi-parietal diameter, crown rump length and fetal abdominal fat content. We are also implementing a similar approach to 3D ultrasound imaging of the placenta.
REFERENCES 1.
University College Dublin. Strategy for Research 2007-2011.
2.
Irish Government Publications. Science, Technology and Innovation 2006-2013
3.
Department of Health and Children. Making Knowledge Work for Health: A Strategy for Health Research
4.
European Commission. Seventh Framework Programme FP7 Cooperation Work Programme: Theme 1 – Health
5.
Fox H. Calcification of the placenta. J Obstet Gynaecol Br Commonw 1964; 17:759
6.
Chitlange SM, Hazari KT, Joshi JV, Shah RK, Mehta AC,. Ultrasonographically observed preterm grade III placenta and perinatal outcome. Int J Gynaecol Obstet 1990; 31: 325-8
7.
Brown HL, Miller JM, Khawli O, Garbert HA. Premature placental calcification in maternal cigarette smokers. Obstet Gynaecol 1988; 71: 914-17
8.
Carroll B. Ultrasonic features of pre-eclampsia. J Clin Ultrasound 1980; 8: 483
9.
Proud J, Grant AM. Third trimester placental grading by untrasonography as a test of fetal wellbeing. Br Med J 1987; 294:1641-4
10. Richardus JH, Graafmans WC, van der Pal-de Bruin KM Amelink-Verburg MP, Verloove-Vanhorick SP, Machenbach JP. A European concerted action investigating the validity of perinatal mortality as an outcome indicator for the quality of antenatal and perinatal care. J Perinat Med 1985; 25: 313-24 11. McKenna D, Tharmaratnam S, Mahsud S, Dornan J. Ultrasonic evidence of placental calcification at 36 weeks' gestation: maternal and fetal outcomes. Acta Obstet Gynecol Scand. 2005 Jan;84(1):7-10 12. Sau A, Seed P, Langford K. Intraobserver and interobserver variation in the sonographic grading of placental maturity. Ultrasound Obstet Gynecol. 2004 Apr;23(4):374-7 13. J.E. Bresenham, "A Linear Algorithm for Incremental Digital Display of Circular Arcs,"Comm ACM, Vol. 20, 1977, pp. 100-106.
Proc. of SPIE Vol. 6920 69200P-7
14. Zimmer Y, Tepper R, Akselrod S (1996) A two-dimensional extension of minimum cross entropy thresholding for the segmentation of ultrasound images. Ultrasound Medic Biol 22:1183–1190 15. Adams R, Bischof L (1994) Seeded region growing. IEEE Trans Pattern Anal Mach Intell 16(6):641–647 16. Pohle R, Toennies K (2001) Segmentation of medical images using adaptive region growing. Proc SPIEMed Im 4322:1337–1346 17. Hao X, Bruce C, Pislaru C, Greenleaf JF (2000) A novel region growing method for segmenting ultrasound images. In: Proceedings of the IEEE international ultrasonics symposium, vol 2, pp 1717–1720 18. Kotropoulos C, Pittas I (2003) Segmentations of ultrasonic images using support vector machines. Pattern Recognit Lett 24:715–727 19. Boukerroui D, Basset O, Guerin N, Baskurt A (1998) Multiresolution texture based adaptive clustering algorithm for breast lesion segmentation. Eur J Ultrasound 8:135–144 20. Thomas JG, Peters RA, Jeanty P (1991) Automatic segmentation of ultrasound images using morphological operators. IEEE Trans Med Im 10:180–186 21. Fan L, Braden GA, Herrington DM (1996) Nonlinear wavelet filter for intracoronary ultrasound images. In: Proceedings of annual meeting on computers in cardiology, pp 41–44 22. Heckman T (1996) Searching for contours. Proc SPIE 2666:223–232 23. Solaiman B, Roux C, Rangayyan RM, Pipelier F, Hillion A (1996) Fuzzy edge evaluation in ultrasound endosonographic images. In: Proceedings of the Canadian conference on electrical and computer engineering, pp 335–338 24. Chen CM, Lu HHS, Hsiao AT (2001) A dual-snake model of high penetrability for ultrasound image boundary extraction. Ultrasound Med Biol 27(12):1651–1665 25. Haenselmann T, Schremmer C, Effelsberg W (2001) Wavelet based semi automatic segmentation of image objects. In: Proceedings of IEEE international conference on signal and image processing, pp 387–392 26. Blake A, Isard M (1999) Active contours: the application of techniques from graphics, vision, control theory and statistics to visual tracking of shapes in motion. Springer, Berlin 27. Paragios N, Deriche R (2002) Geodesic active regions and level set methods for supervised texture segmentation. Int J Comput Vis 46(3):223–247 28. McInerney T, Terzopoulos D (1996) Deformable models in medical images analysis: a survey. Med Im Anal 1(2):91–108 29. Suri JS, Liu K, Laxminarayan SN, Zeng X, Reden L (2002) Shape recovery algorithms using level sets in 2D/3-D medical imagery: a state-of-the-art review. IEEE Trans Inf Tech Biomed 6(1):8–28 30. I. Dindoyal, T. Lambrou, J. Deng, C.F. Ruff, A.D. Linney, C.H. Rodeck and A. Todd-Pokropek. “Level Set Segmentation of the Fetal Heart.” Lecture Notes in Computer Science (2005), 3502:123-132. 31. Iakovidis D, Savelonas M, Karkanis S, Maroulis D. “Genetically optimized level set approach to segmentation of thyroid ultrasound images.” (2007) Applied Intelligence. 27(3):193-203
Proc. of SPIE Vol. 6920 69200P-8
