Uporaba umetne inteligence v ultrazvočni diagnostiki - sistematični pregled literature : diplomsko delo

Kosi, Alen (author), Kegl, Nik (author), Žibert, Janez (mentor), Medič, Mojca (reviewer)

Uvod: Ultrazvočna diagnostika temelji na detekciji in prikazu akustične energije, ki se odbije od različnih struktur v telesu. Pridobljene podatke uporabimo za prikaz visoko resolucijske sivo-tonske sivine telesa. Lahko prikažemo tudi pretok krvi v telesu. Za izvajanje ultrazvočne preiskave uporabljamo ultrazvočno sondo. Uporabljamo različne tehnike zajema podatkov, ki so A-mode, B-mode, M-mode in Dopplerjev učinek. Pridobljene podatke procesiramo s centralno procesirno enoto, ki podatke interpretira in prikaže na monitorju. Umetna inteligenca je v radiologiji napredek, ki raste iz dneva v dan in je že pokazalo obetavne sposobnosti pri razlikovanju, prepoznavanju in določanju patologije. Umetna inteligenca pomaga zdravstvenim delavcem pri opravilih, ki zahtevajo veliko časa in podrobno analizo večjega števila podatkov. Namen: Namen diplomskega dela je preučiti, kako se umetna inteligenca uporablja v ultrazvočni diagnostiki. Specifično na področjih ultrazvočne preiskave jeter, ščitnice, dojk in ginekološkega področja. Metode dela: Uporabili smo metodo sistematičnega pregleda literature. Literaturo smo iskali iz podatkovnih baz Google Scholar, ScienceDirect in PubMed. Članke smo našli z uporabo vključitvenih in izključitvenih kriterijev. Rezultati: Po natančnem pregledu izbrane literature smo predstavili uporabljene modele umetne inteligence in zakaj so jih uporabili. Umetna inteligenca se v ultrazvočni diagnostiki uporablja za diagnostiko, detekcijo in klasifikacijo patologij. Pri ultrazvočni preiskavi na področju jeter se je najboljše izkazal model umetne inteligence imenovan ResNet50 za detekcijo in diagnosticiranje jetrnih lezij. Model LivGuard se je uporabil kot klasifikator ciroze. Za področje ultrazvočne preiskave ščitnice se je za najboljši model umetne inteligence za detekcijo ščitničnih nodulov izkazal model imenovan ResNet-34. Za diagnosticiranje pa sta najboljša modela Xception in CNN-VGG, ki sta dosegla primerljive rezultate. Najboljši modeli za diagnosticiranje med benignimi in malignimi lezijami v dojkah so bili CNNE1, CNNE2, CNNE3 in Bu-CAD sistem umetne inteligence. Ti modeli so dosegli podobne rezultate. Najboljši klasifikator glede na benigne ali maligne lezije jajčnikov je bil Ovry-Dx2. DLR model se je uporabil kot pomoč pri diagnostiki raka endometrija. Model YOLOv8 pa se je uporabil kot detektor za rak endometrija. Razprava in zaključek: Umetna inteligenca predstavlja velik potencial v ultrazvočni diagnostiki. Uporablja se za diagnosticiranje, lokalizacijo in klasifikacijo različnih patologij. Čeprav modeli umetne inteligence dosegajo primerljive, nekateri tudi boljše rezultate kot izkušeni radiologi, še niso dovolj testirani na večjem številu podatkov, da bi lahko bili modeli samostojno uporabljeni v diagnostiki. Iz člankov smo razbrali, da je umetna inteligenca na dobri poti, da bi bila ultrazvočna preiskava avtomatizirana, vendar še ni dovolj razvita, da bi nadomestila delo izkušenega radiologa.

• Language: Slovenian
• Keywords: diplomska dela, radiološka tehnologija, umetna inteligenca, AI, ultrazvok, ščitnica, dojke, ginekološkega področja, jetra, ultrasonografija
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: ZF - Faculty of Health Sciences
• Place of publishing: Ljubljana
• Publisher: [A. Kosi : N. Kegl]
• Year: 2026
• Number of pages: 46 str.
• PID: 20.500.12556/RUL-188011
• UDC: 616-07
• COBISS.SI-ID: 291415043
• Publication date in RUL: 17.09.2026

Secondary language

• Language: English
• Title: Use of artificial intelligence in ultrasound diagnostics - systematic review of literature : diploma work

Abstract

Introduction: Ultrasound diagnostics is based on the emission, detection and to display acustic energy that echoes fromdifferent structures in the body. The acquired data is used to display high resolution gray-scale images of the body. We can also show the blod flow. For the procedure to take place we use an ultrasonic probe. To capture the image we use different methods such as A-mode, B-mode, M-mode and Doppler effect imaging technique. We use a central procesing uni to interpret the acquired data in display them on the screen. Artificial intelligence is a growing phenomenon in radiology and it has shown promising results in detecting, differentiating and determening pathologies. Artificial intelligence helps health workers with tasks that require a lot of time and detailed analysis of high volume of data. Purpose: Purpose of the diploma work is to study how artificial intelligence is used in ultrasound diagnostics. Specifically in the ultrasound examinations of the liver, thyroid, breasts and in the gynecological area. Methods: We used a systematic review of literature method. The literature was acquired from the following data bases Google Scholar, ScienceDirect and PubMed. The articles were found using including and excluding criteria. Results: After a thorough analysis of the selected literature we presented the artificial inteligence model that was used in the article and what was its purpose. Artificial inteligence is used in ultrasound diagnostics as a detection, diagnostic and classification tool for different pathologies. The best artificial inteligence model for detection and diagnosing focal liver lessions turned out to be ResNet50. The model LivGuard was used for cirhosis classification For the ultrasound diagnostics of the thyroid, the best model used was ResNet-34 for detection of thyroid nodules. The best AI models used for diagnosing thyroid nodules were Xception and CNN-VGG. These models showed similar results. The best diagnostic model for diferentiating between benign and malignant breast lession were CNNE1, CNNE2, CNNE3 and BU-CAD with similar results. The best classificator between benign and malignant ovary lessions was Ovry-Dx2. A DLR model was used for helping with the diagnosis of endometrium cancer and the model YOLOv8 was used for detecting endometrium cancer on ultrasound images. Discussion and conclusion: Artifical intelligence poses great potential in ultrasound diagnostics. Currently it is mainly used for classification, detection and diagnosing of different pathologies. Even though artificial intelligence models achive good results, some even better than experienced radiologists, they are not yet tested enough on bigger trials with more data to be used as an independently in diagnostics. From the articles we can conclude that artificial inteligence is on a good way to make the ultrasound procedure fully automatic, but it is not yet ready to replace an experienced radiologist.

• Keywords: diploma theses, radiologic technology, artificial inteligence, AI, ultrasound, thyroid, breast, gynecological area, liver, ultrasonography