Uporaba algoritmov umetne inteligence pri detekciji pljučnih nodulov : magistrsko delo

Jošt, Jaka (author), Žibert, Janez (mentor), Viltužnik, Rebeka (comentor), Fošnarič, Miha (reviewer)

Uvod: Radiologija predstavlja vodilno vlogo pri raziskovanju medicinske digitalne tehnologije in uporabe umetne inteligence (UI). V zadnjih letih so se uveljavili algoritmi umetne inteligence, ki lahko radiologom pomagajo pri njihovem vsakdanjem delu. Reševanje določenih zahtevnih primerov pljučnega raka so lahko za radiologe zahtevne, kar pa je za uporabo algoritmov UI lažje in nezahtevno. Namen: Namen magistrskega dela je bralcu predstaviti področje umetne inteligence s poudarkom na detekciji pljučnih nodulov in na podlagi serije primerov ugotoviti kako lahko orodja UI pripomorejo k olajšanju dela radioloških inženirjev, radiologov in omogočijo hitrejšo obravnavo bolnikov. Metode dela: Raziskavo smo izvedli v dveh delih. V prvem delu smo uporabili deskriptivno metodo dela. Pri tem smo uporabili različne podatkovne baze kot so Medline, PubMed... . V drugem delu pa smo s pomočjo odprtokodnih programov izvedli študijo primera, s katero smo ugotovili učinkovitost orodij umetne inteligence. V študiji smo uporabili 20 primerov, katere smo pridobili v podatkovni bazi Kliničnega inštituta za radiologijo, Univerzitetnega kliničnega centra Ljubljana in podatkovne baze LUNA16. Rezultati: V sklopu magistrskega dela smo ugotovili, da model Lung Nodule CT Detection, ki je del MONAI Label, v primeru slik iz vsakdanjega kliničnega okolja odkrije 14 resnično pozitivnih nodulov in 53 lažno pozitivnih nodulov. V magistrskem delu smo želeli preveriti tudi, kako dober je model Lung Nodule CT Detection v primeru slik iz podatkovne baze LUNA16, na kateri je bil model tudi naučen. Ugotovili smo, da model zazna 18 resnično pozitivnih nodulov in 137 lažno pozitivnih nodulov. Tekom raziskave smo ugotovili, da se s parametroma score_tresh in nms_tresh lahko bistveno izboljša razmerje lažno pozitivnih rezultatov. Razprava in zaključek: V sklopu našega magistrskega dela smo ugotovili, da model Lung Nodule CT Detection po predhodni označitvi ROI dodatno označi pljučni nodul in ga pravilno detektira. Model ima nekaj pomanjkljivosti, saj nam poda tudi nekaj FP rezultatov, ki pa jih lahko v prihodnje še zmanjšamo s parametroma score_tresh in nms_tresh.

• Language: Slovenian
• Keywords: magistrska dela, radiološka tehnologija, umetna inteligenca, CAD, računalniška tomografija, pljučni nodul, segmentacija, avtomatska detekcija
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: ZF - Faculty of Health Sciences
• Place of publishing: Ljubljana
• Publisher: [J. Jošt]
• Year: 2025
• Number of pages: 79 str., [3] str. pril.
• PID: 20.500.12556/RUL-174672
• UDC: 616-07
• COBISS.SI-ID: 252353027
• Publication date in RUL: 08.10.2025

Secondary language

• Language: English
• Title: Application of artificial intelligence algorithms in pulmonary nodule detection : master thesis

Abstract

Introduction: Radiology is at the forefront of research into medical digital technology and the use of AI (Artificial Intelligence). In recent years, AI algorithms have emerged that can help radiologists in their daily work. Dealing with certain complex cases of lung cancer can be challenging for radiologists, which in turn is easier and undemanding for the use of AI algorithms. Purpose: The aim of this master thesis is to introduce the field of artificial intelligence to the reader, with a focus on pulmonary nodule detection, and to use a series of case studies to see how AI tools can help to facilitate the work of radiographers, radiologists and enable faster treatment of patients. Methods: The survey was conducted in two parts. In the first part, we used a descriptive method. We used different databases such as Medline, PubMed... . In the second part, we conducted a case study using open source software to determine the effectiveness of AI tools. In the study, we used 20 cases obtained from the database of the Clinical Institute of Radiology, University Clinical Centre Ljubljana and from the freely available LUNA database16. Results: As part of our master thesis, we found that the Lung Nodule CT Detection model, which is part of the MONAI Label, detects 14 true positive nodules and 53 false positive nodules images from everyday clinical settings. In the master thesis we also wanted to check how good the Lung Nodule CT Detection model is in the case of images from the LUNA16 database on which the model was also trained. We found that the model detects 18 true positive nodules, 137 false positive nodules. In the course of our research, we found that the score_tresh and nms_tresh parameters can significantly improve the ratio of FP scores. Discussion and conclusion: As part of our master thesis, we found that the Lung Nodule CT Detection model detects a lung nodule after pre-labelling the ROI and detects it correctly. The model has some shortcomings as it also gives us some FP score, which can be further reduced in the future by using score_tresh and nms_tresh parameters.

• Keywords: master's theses, radiologic technology, artificial intelligence, CAD, computed tomography, pulmonary nodule, segmentation, automated detection