Ocena učinkovitosti segmentacije in detekcije pri računalniško vodenem poteku diagnoze na mamografskih slikah

• Main Authors: Cvar, Petja, Milič, Jaro
• Other Authors: Žibert, Janez
• Format: Bachelor Thesis
• Language: Slovenian
• Published: 2019
• Subjects: računalniško voden potek diagnoze; mamografija; segmentacija; upragovljanje; rast področij; rojenje; strojno učenje; mamografska baza DDSM; computer aided detection; segmentation; mammography; thresholding; region growth; clustering; machine learning; DDSM database; Baza; sami
• Online Access: https://repozitorij.uni-lj.si/IzpisGradiva.php?id=111604; https://repozitorij.uni-lj.si/Dokument.php?id=123482&dn=; https://plus.si.cobiss.net/opac7/bib/5705067?lang=sl

Uvod: Mamografija se je v zadnjih letih razširila kot primarna diagnostična preiskava za detekcijo bolezni dojk, predvsem kot metoda izbora presejalnih programov. Pri sami interpretaciji mamogramov, pravočasnem odkrivanju karcinomov in razlikovanju med benignimi in malignimi tumorskimi masami so v veliko pomoč orodja računalniško vodene diagnoze. Delijo se na Computer Aided Detection (CADe) in Computer Aided Diagnosis (CACx). Omenjena orodja se pri svojem delovanju poslužujejo metod strojnega učenja. Razdelitev slike v homogene regije tekstur je eden prvih korakov pri razumevanju, analizi in poglobljenem vpogledu v dano sliko. Med CAD metode med drugim spadajo tudi segmentacijski algoritmi pri obdelavi digitalnih slik. V nalogi smo osredotočeni na segmentacije z upragovljanjem, segmentacije z določanjem območij in segmentacije z učenjem. Namen: Namen diplomske naloge je ugotoviti, katera izmed izbranih oblik segmentacije v sklopu metod računalniško vodene detekcije najbolj ustrezno segmentira sliko pri izbrani bazi mamografskih slik. Metode dela: Pri pisanju diplomske naloge je bila uporabljena metoda deskripcije pri opisovanju pojmov in metoda kompilacije pri uporabi izpiskov, navedb in citatov drugih avtorjev. Nato je na podlagi knjižnih virov opravljen kvalitativni pregled gradiva z namenom primerjave načinov segmentacije in detekcije interesnih področij na mamografskih slikah izbrani bazi mamografskih slik. Rezultati: Primerjane so študije, ki testirajo svoje metode segmentacije in detekcije na javno dostopni bazi mamografskih slik Digital Database For Screening Mammography (DDSM). Opravljen je pregled rezultatov v obliki tabele in ovrednoteni so posledični izsledki. Komentirali smo prednosti in slabosti različnih metod in predlagali najučinkovitejšo. Razprava in zaključek: Zaključimo lahko, da metode, predlagane s strani pregledanih študij zadovoljivo interpretirajo sliko pri izbrani bazi mamografskih slik, a ne dosegajo enako konsistentnih rezultatov kot zdravniki specialisti. Na podlagi pregledanih raziskav lahko sklepamo, da višje rezultate dosegajo segmentacije, ki uporabljajo pri svojem delovanju strojno učenje in proces rojenja. Težavo pri doseganju ponovljivih in primerljivih rezultatov študij predstavlja uporaba različnih mamografskih slik za analizo in adaptacija nevronskih mrež s strani raziskovalcev. Preden se lahko CADe (angl. Computer Aided Detection) metode uvrstijo med komplementarne tehnike pri diagnostiki mamografskih slik v sami klinični praksi, je potreben nadaljnji razvoj področja in konsistentno doseganje zadovoljivih rezultatov, predvsem visokih vrednosti senzitivnosti, točnosti in AUC. Introduction: Mammography has become the number one detection method of breast cancer in the recent years, especially through various preliminary screening programs. Mammogram analysis through computer aided detection has been established as a big aid to radiologists in early cancer detection rates. Computer aided detection (CADe) represents a segment of Computer aided diagnosis (CAD), both of which employ the methods of machine learning in their workings. One of prerequisites for efficient detection of tumor masses is adequate segmentation of presented breast tissue. This work is focused mainly on threshhold based segmentation, region based segmentation and segmentation based on learning. Purpose: We intended to establish the efficiency of segmentation methods and positive detection rates used in modern computer aided detection processes. Methods: A descriptive method was used to explain the basic concepts of segmentation and detection of cancer tissue in CADe methods through extensive study of available material on current research of the field in question. The results were presented in a qualitative manner with a commentary on efficiency and viability of methods used. Results: Studies, that tested their segmentation and CADe methods on the publicly available database Digital Datbase for Screening Mammography (DDSM), were reviewed. We compared selected studies from the field of computer aided detection and assesed their efficiency in breast tissue segmentation and positive detection rates of cancer mass. Discussion and conclusion: It was concluded that CADe methods adequately segment and detect cancer tissue in mammograms, but do not yet reach the efficiency of trained radiologists. It is evident that methods employing machine learning algorithms and clustering segmentation tend to produce better overall results than the rest of reviewed methods. The studied sources suggest there is a lack of uniform, publicly accessible mammogram databases that could be used to further research the field with practically comparable results. As such, CADe methods and the segmentation processes involved show promise in the future of automatic interpretation of mammography screening.