Classification of X-ray images using deep learning

NELA, ARDIT (author), Skočaj, Danijel (mentor)

As technology continues to advance, many different fields are exploring the benefits of machine learning, and medicine is no exception. The application of machine learning in medicine has the potential to revolutionize the way we approach patient care and disease diagnosis. One area in which machine learning has shown particular promise is the analysis of medical imaging. For example, chest X-rays are a common diagnostic tool used to detect a variety of lung diseases, including pneumonia, hernia, and cardiomegaly. However, accurately interpreting X-ray images requires significant expertise and experience on the part of a medical professional. In the thesis, we discuss the problem of automatic classification of X-ray images. We use data collected and labeled by the NIH to train a state-of-the-art machine learning model based on deep neural networks to detect any issues in the X-ray scan. We compared different architectures and determined that ResNet101 and Transformers performed the best. Using these models alongside transfer learning resulted in mean AUC scores of approximately 0.795.

• Language: English
• Keywords: neural networks, deep learning, computer vision, image classification, medical imaging
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-152708
• COBISS.SI-ID: 169373955
• Publication date in RUL: 04.12.2023

Secondary language

• Language: Slovenian
• Title: Klasifikacija rentgenskih slik z uporabo globokega učenja

Abstract

Ker tehnologija še naprej napreduje, različna področja raziskujejo prednosti strojnega učenja, in medicina ni nobena izjema. Uporaba strojnega učenja v medicini ima potencial, da bistveno spremeni način, kako pristopamo k oskrbi pacientov in diagnozi bolezni. Eno izmed področj na katerem je strojno učenje pokazalo posebno obetavne rezultate, je analiza medicinskih slik. Na primer, rentgenski posnetki prsnega koša so pogosto diagnostično orodje, ki se uporablja za odkrivanje različnih pljučnih bolezni, vključno s pljučnico, kilo in kardiomegalijo. Vendar pa za natančno interpretacijo rentgenskih slik potrebujemo strokovnost in izkušnje s strani zdravstvenih strokovnjakov. V diplomskem delu obravnavamo problem avtomatske klasifikacije rentgenskih slik. Za učenje modela, ki temelji na globokih nevronskih mrežah za zaznavanje morebitnih nepravilnosti na rentgenskem posnetku, uporabljamo podatke, zbrane in označene s strani NIH. Primerjali smo različne arhitekture in ugotovili, da se ResNet101 in Transformers najbolje izkažeta. Z uporabo teh dveh modelov smo dosegli povprečne AUC ocene približno 0,795.

• Keywords: nevronske mreže, globoko učenje, računalniški vid, klasifikacija slik, medecinske slike