CC BY
15
Научни трудове на Съюза на учените в България-Пловдив, Серия Г. Медицина, фармация и дентална медицина т. XIX. ISSN 1311-9427 юни 2016. Scientific works of the Union of Scientists in Bulgaria-Plovdiv, series G. Medicine, Pharmacy and Dental medicine, Vol. XIX, ISSN 1311-9427 Medicine and Dental medicine June 2016.
МАТЕМАТИЧЕСКИ МОДЕЛИ НА РАКОВИ ОБРАЗУВАНИЯ НА ГЪРДАТА ОТ РЕАЛНИ ИЗОБРАЖЕНИЯ ОТ КОМПЮТЪР ТОМОГРАФ: ПРЕДВАРИТЕЛНИ РЕЗУЛТАТИ Николай Дуков1, Фирган Ферадов1, Кристина Близнакова1, Елица Енчева2, Яна Глухчева2, Даниел Буляшки3, Радослав Радев3 1Факултет по Изчислителна Техника и Автоматизация, ТУ - Варна 2Клиника по Лъчелечение, УМБАЛ „Св Марина", Варна 3Клиниката по гръдна хирургия, УМБАЛ „Св Марина", Варна
COMPUTATIONAL BREAST CANCER MODELS CREATED FROM
PATIENT SPECIFIC CT IMAGES: PERLIMINARY RESULTS Nikolay Dukov1, Firgan Feradov1, Kristina Bliznakova1, Elica Encheva2, Yana Gluhcheva2, Daniel Bulyashki3, Radoslav Radev3 1Faculty of Automation and Computing, Technical University - Varna 2 Radiotherapy Department, "St. Marina" University Hospital, Varna 3Clinic of thoracic surgery UMHAT "St. Marina" University Hospital,
Varna
Abstract: Breast cancer remains the most common cause of death for women below seventy years of age. Although, screening nowadays is a common practice the standard tools for such procedure in some cases of breast cancers are not as efficient as desired. New approaches are constantly being developed to detect and diagnose the cancerous formations as earlier as possible. These new techniques require extensive optimization of parameters which is best performed with computer-based models. Our main objective is the creation of comprehensive breast cancer computer database for the purposes of developing, testing and optimizing new x-ray imaging techniques. This paper reports on a semi-automatic approach for segmentation of cancerous tissue extracted from patient specific CT datasets and the creation of solid breast cancer models. Key words: breast cancer, computer-based breast cancer models, patient CT images, segmentation
INTRODUCTION
Breast cancer is the leading cause of death for women below seventy years of age. In Europe, one in ten women will develop breast cancer in her lifetime. Early diagnosis is recognized as a critical factor that improves the chance of survival. Nowadays, the standard tool for breast cancer screening is the digital mammography. However, screening and diagnosing cancers hidden in breast dense parenchyma with digital mammography remains a challenging task. Approximately 10% to 20% of palpable breast cancers are still not visible in planar mammograms due to insufficient soft tissue contrast and the effect of overlying structures in this modality [1]. New approaches to breast imaging are continuously under development. Breast tomosynthesis is one such approach. This technique is able to produce three-dimensional (3D) structural information of the breast
in which the influence; of overlapping tissues is greatly reduced and the accuracy of cancer detection is encrepced, particuChriy foc C^e^i^^^l mas^s iP]. ehnoPihi e^j^j^romig ig d^e braast C!T [3]. The intsndrntian ofigesn; la cin^ques foc routine screening mammo graphy rrpnieas c^rinm^s^i^^itp hf pa^;^m^(tte rs and algorithms, which process is best carried out by using computer models.
Our iong term goal is to create a computer database nitC breast panphe mrehis foe eCh hPehdsss of ssudees of new x-ray maging techniqpss ahhiihe to Orease. TCh main coneriOption to this rim is to develop segmentation algorithms foe breast massss nith iereepiae saahss feom hatisne shspiilp d^t^ia. Ties paper presents a reim-automatic algorithm foe segmentation of crncseops tissprns feom CT patient imaess.
CT IMAGES
Two sehs of abdoimnal CT images were obtainhe at the Umvsesiny Hoshieri "Scine Marina" in Viffna, usmg SOMATOM (Siemens). The acquisition was maeh by utilizmg a senneaee heotocoi that provides images of s^^e 51e x \6 bits grey (evel resolption. Foe convsetsncs fprtter on in tWs p^j^^r the two sets of CT (mages are refheree as S1 rne Se. Thh thtcknsss of tte siicss .s emm for both cqses, while tice numbhe of ilices Is 177 ene 134 foe S1 mee S e, res^ctrvrnfy. Thh imaess are with square pixel size of 0.9766 mm for S1 and 1.2695 mm for S2.
Fig. 1. Biock eingfnm of thh rleoiteam foe Orerst prepse ssemseereioe.
The findings on the CT images were classified as malignant tumors by an experienced radiologist who works in this area. The two cancer formations are characterized with different size and location within the breast. Specifically, twenty-two slices were found to contain information about the malignant tumor formation in S1 and thirty CT images in S2. All images were anonymized prior processing.
SEGMENTATION ALGORITHM
The obtained DICOM images were processed to leave only the information about the breast cancer by applying an in-house developed algorithm for semi-automatic segmentation. The main stages of the proposed algorithm are shown in Fig.1, where CT images correspond to slices taken at different location in the patient CT scan.
Initially, a region of interest (ROI) is selected for each of the slices, in order to reduce the size of the segmented area and therefore to improve the efficiency of the proposed segmentation algorithm. From the selected ROI, a binary mask is generated, where the values of the pixels inside the ROI are set to ones and values of the pixels outside the ROI to zeros. The generated masks are then applied to the slices from the CT sets which results into the desired area subjected to segmentation. Consequently, the preprocessed CT images were subjected to image thresholding. The threshold value was chosen based on an in-house developed adaptive thresholding algorithm. Thresholding is a necessary process, which results in rough segmentation of the cancerous tissue. Further on, the CT images were converted to binary images on a pixel by pixel basis, as the values of the pixels which do not belong to the breast cancer were set to zero, while these which belong to the cancerous tissue were set to one. Although S1 responded well to this technique the case of S2 proved more challenging. Due to the increased image noise in the images from set S2, the segmented final cancer volume did not appear solid. This issue opted for a multiple iterations of the morphological operation dilation followed by an erosion operation, both performed with a small sized diamond-shaped structuring element. These operations correct the existing imperfections caused by the noise while maintaining the shape of the segmented objects (Fig.2).
nn
Fig. 2. Originally segmented image, dilated image, eroded image (from left to right).
However, often in the different slices, for both image sets, objects not containing cancer information were segmented to belong to cancerous tissue. For convenience, these objects are referred as artefacts. To lower the need of interactively removing artefacts, a series of image processing operations were applied to the already segmented images (in binary form).
In order to remove smaller artefacts morphological area opening was applied to each slice, where objects with a given maximum number of pixels were excluded. In the case оf S1, objects with fewer than 50 pixels seemed appropriate for removal. Nonetheless, attention should be paid to the parameters of the area opening operation as cancer information could be lost due to its size in some slices. Such is the case of S2, where in order to preserve information about the malignant formation, objects with no more than 5 pixels were removed.
Finally, morphological erosion with a diamond-shaped structuring element was performed on each slice followed again by a morphological area opening to eliminate any remaining artefacts.
The segmented volumes are then stored in a 3D matrix. Post-processing of the 3D models includes the application of different 3D filters which results in smoothing the edges of the voxelized computational cancer model (Fig.3), as well as interpolation techniques to scale properly the final model.
UO l0U \20 2S0
Fig. 3. 3D models of the segmented malignant tumor formations from S1 and S2.
EVALUATION
An experienced radiologist evaluated the realism of the new computational breast cancer models. The created computational based breast cancer models were stored in a database with detail information for the size, the origin and the description given by the pathologists. The subjective evaluation showed satisfactory realism of the generated 3D computational models.
CONCLUSION
This paper presented preliminary results of the development and application of an algorithm for segmentation of breast cancers from patient specific CT images. Next is to create more computational models of breast cancers, which will be used as a base for the modelling of realistic mathematical models of breast cancers. The achieved results, which apart from applications in 3D breast imaging research, are also encouraging for educational and training purposes.
ACKNOWLEDGEMENTS
This research has been supported by the MaXIMA project: Three dimensional breast cancer Models for X-ray IMAging research, from the H2020-TWINN-2015 (Project Number: 692097).
REFERENCES
[1]. Schulz-Wendtland R., Fuchsjäger M., Wacker T., Hermann K., "Digital mammography: an update" (2009), Eur J Radiol. 72 (2), 258-265.
[2]. Sechopoulos, I., "A review of breast tomosynthesis. Part I. The image acquisition process" (2013), Medical Physics, 40 (1), art. no. 014301.
[3]. Kalender W.A., Kolditz D., Steiding C., Ruth V., Lück F., Rößler A.-C., Wenkel E., "Technical feasibility proof for high-resolution low-dose photon-counting CT of the breast" (2016), European Radiology, pp. 1-6. Article in Press.
