Primjena konvolucijskih neuronskih mreža u analizi rendgenskih slika prsnog koša

master's thesis

• Title (english): Application of Convolutional Neural Networks in the Analysis of Chest X-ray Images
• Author: Domagoj Razum
• Mentor: Igor Lacković (mentor)
• Committee member: Igor Lacković (predsjednik povjerenstva)
• Committee member: Ratko Magjarević (član povjerenstva)
• Committee member: Dinko Oletić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-09-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Rad se fokusira na automatsku detekciju patoloških stanja prsnog koša iz rendgenskih slika primjenom metoda dubokog učenja. Korištene su konvolucijske neuronske mreže i tehnika prijenosnog učenja u svrhu klasifikacije ChestX-Ray14 skupa podataka. Nadalje, detaljno je opisan skup podataka s 14 klasa bolesti koji je javno dostupan na službenim stranicama National Institutes of Health. Razložen je postupak gradnje arhitekture modela, učitavanja i obrade podataka, nakon čega slijedi pregled procesa treniranja i testiranja modela. Korišteni su programski jezik Python zajedno s Tensorflow bibliotekom. Arhitektura klasifikacijskog modela koristi metode prijenosnog učenja s modelom EfficientNet-B0. Zbog memorijskih zahtjeva treniranje je izvršeno na AWS servisu SageMaker. Kao metrike rezultata korišteni su ROC i AUROC. Pokriveni su i problemi koji se javljaju prilikom rada na kompleksnim i resursno zahtjevnim modelima, kao i problemi prilikom rada s nebalansiranim skupom podataka. Prosječne dobivene vrijednosti AUROC svih patoloških stanja su 0,7766 i 0,7711 za različite vrijednosti epoha. Na kraju su opisana potencijalna poboljšanja programskog rješenja u smislu odabira parametara za daljnja istraživanja.

Abstract (english)

The focus of the work is on the automatic detection of chest pathologies from x-ray images by applying deep learning methods. Convolutional neural networks and transfer learning technique have been used for the purpose of the classifying ChestX-ray14 dataset. Furthermore, dataset with 14 disease classes which is publicly available on the National Institutes of Health site has been described in details. The procedure of building a model architecture, data loading and processing is explained, after which comes an overview of model training and testing process. Programming language Python alongside Tensorflow library have been used. Classification model architecture uses transfer learning method with EfficientNet-B0 model. Due to memory requirements, training was performed on AWS service SageMaker. ROC and AUROC have been used as metrics. The problems that arise when working on complex and resource intensive models have been addressed, as well as problems when working with unbalanced dataset. The average obtained values of AUROC of all pathological conditions were 0.7766 and 0.7711 for different epochs. Finally, potential improvements of the software solution are described.

• Keywords: konvolucijske neuronske mreže; prijenosno učenje; rendgenske slike; klasifikacija; EfficientNet-B0; Adam optimizator; ROC; AUROC
• Keywords (english): convolutional neural networks; transfer learning; x-ray images; classification; EfficientNet-B0; Adam optimizer; ROC; AUROC
• Language: croatian
• URN:NBN: urn:nbn:hr:168:071616
• Study programme: Title: Information and Communication Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije (magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 13:41:38