Muštra, Mario

Segmentacija mamograma i detekcija mikrokalcifikacija adaptivnim poboljšanjem kontrasta : doktorski rad / Mario Muštra ; mentor Mislav Grgić - Zagreb: M. Muštra ; Fakultet elektrotehnike i računarstva, 2013. - 167 str.: ilustr. dijelom u bojama; 30 cm + CD

Bibliografija: str. 157-163. - Sažetak na hrv. i eng. jeziku. - Životopis [uključuje i popis radova autora]

SAŽETAK: Mamografija je neinvazivna metoda pregleda dojki koja koristi rendgensko zračenje u svrhu stvaranja slike tkiva dojke. Računalno potpomognuta dijagnostika dobiva sve veći značaj u medicini razvojem računalne tehnologije. Mogućnost automatske detekcije objekata u digitalnoj slici i donošenja automatske dijagnoze na osnovi izdvojenih značajki, važan su alat koji se sve više koristi u modernoj dijagnostici. U ovom radu predložena je metoda za segmentaciju tkiva dojke od pozadine i izdvajanje prsnog mišića iz mamograma koji prikazuju postraničnu projekciju. Segmentacija tkiva dojke od pozadine i uklanjanje objekta izvan tkiva dojke postignuti su kombiniranom uporabom morfoloških operatora i lokalnim poboljšanjem kontrasta rubnih dijelova tkiva dojke u slikama. Izdvajanje prsnog mišića postignuto je određivanjem koeficijenata polinoma trećeg stupnja pomoću detektiranih vidljivih točaka na rubu prsnog mišića. Ovaj pristup za određivanje ruba prsnog mišića izabran je iz razloga što daje bolje rezultate u slučajevima slabije vidljivosti ili potpune nevidljivosti dijela prsnog mišića u odnosu na ostalo tkivo dojke. Određivanje gustoće dojke temelji se na izdvajanju teksturnih značajki iz područja žljezdanog diska i klasifikacijom u neku od kategorija prema gustoći. Detekcija mikrokalcifikacija predstavlja problem zbog nejednolike teksture tkiva dojke i različitih uvjeta detekcije u ovisnosti o gustoći pojedine dojke. Predložena metoda za detekciju područja koja sadrže mikrokalcifikacije uključuje znanje o gustoći dojke kao parametra pri automatskoj detekciji. Učinkovitost predloženih metoda ispitana je na različitim bazama mamografskih slika, kako bi se učinkovitost pojedine metode bolje verificirala. - KLJUČNE RIJEČI: mamografija, segmentacija, lokalno poboljšanje kontrasta, izdvajanje prsnog mišića, detekcija mikrokalcifikacija, baza mamografskih slika SUMMARY: Mammography is a non invasive breast screening method which uses X-ray radiation to capture breast tissue images. Computer aided diagnosis gains larger influence in medical science with the development of new information technologies. Possibility of automatic object detection in digital images and automatic diagnosis process based on extracted features are important tools used more and more in modern diagnostics. In this thesis a method for automatic breast tissue segmentation from the background and pectoral muscle extraction from Medio-lateral oblique view mammograms have been presented. The automatic breast density classification method is based on extraction of textural features from the fibroglandular discs and classification into one of density categories. The automatic detection of microcalcifications presents a problem because of uneven breast tissue texture and different detection conditions according to the actual breast density. The proposed method for detection of areas which contain microcalcifications has been proved on two different mammography databases in order to better verify its efficiency. This thesis consists of seven chapters. The first chapter brings a short introduction and gives an overview of mammography, X-ray imaging, mammography databases and developed algorithms. The second chapter gives a description of X-ray imaging techniques and their applications. Analog and digital X-ray detectors used for capturing of mammograms have been described and for each different detector its advantages and disadvantages have been shown. The last section of the second chapter brings comparison of analog and digital mammography, both in image quality and ease of use as well as in glandular radiation doses. In the third chapter an overview of publicly available mammography databases, MIAS and DDSM, is given. This chapter also gives basic information about morphological image processing techniques and gives an overview of the preprocessing techniques which need to be performed in order to make a captured mammogram ready for usage in automatic detection and diagnosis algorithms. In this chapter a method for the extraction of manually segmented breast tissue and pectoral muscle masks has been described. This method uses the Hough transform to align images in the same manner by detecting straight lines which denote the breast tissue border and the combination of morphological operations to extract the manually drawn segmentation lines. The fourth chapter deals with the development of the method for automatic breast skin-line and pectoral muscle segmentation. The proposed method for automatic breast skin-line segmentation presents a combination of morphological operations used to extract the breast tissue mask. On the detected area where the actual skin-line is situated a contrast enhancement method is applied for every segment of the extracted area in order to achieve a precise segmentation of the low-contrast edges. The method for the pectoral muscle segmentation uses a similar approach but has some different steps because of different properties of segmentation objects and backgrounds. For the pectoral muscle segmentation a polynomial fitting has been proposed in order to achieve good segmentation in cases when the intensity of the pectoral muscle, or part of the pectoral muscle, is similar or the same as the rest of the breast tissue. This chapter also brings results of the proposed automatic segmentation methods with the manual segmentation from an expert radiologist. In the fifth chapter a method for automatic breast density classification has been presented. The proposed method uses a classification into different number of categories based on extracted textural features from fibroglandular discs. From the fibroglandular disc first and second order statistical features have been extracted and using different feature selection algorithms the features which give the best classification results have been selected. Classification has been carried out using k-NN and Bayesian classifier and results show that there is no significant difference in classification accuracy when varying choice of classifiers used with the extracted features. As expected, classification accuracy depends significantly only on number of categories in which instances are being classified. The sixth chapter presents a method for the detection of regions which contain microcalcifications. The proposed method uses the Discrete Wavelet Transform (DWT) in order to filter out components of the spatial frequencies which do not correspond to microcalcifications. From the reconstructed images, after DWT filtering, a contrast enhancement method was applied in order to boost the visibility of microcalcifications. By knowing the breast density as an input parameter, this method applies different wavelet decomposition level and threshold settings in order to achieve better segmentation of microcalcifications and maintain as low as possible false positive rate. - KEYWORDS: mammography, segmentation, local contrast enhancement, pectoral muscle extraction, detection of microcalcifications, mammography image database

004.932.72'1 618.19‑073.7‑71

Središnja knjižnica Fakulteta elektrotehnike i računarstva, Unska 3, 10000 Zagreb
tel +385 1 6129 886 | fax +385 1 6129 888 | ferlib@fer.hr