Duboko učenje za segmentaciju srca

master's thesis

• Title (english): Deep Learning for Whole Heart Segmentation
• Author: Marko Meštrović
• Mentor: Sven Lončarić (mentor)
• Committee member: Sven Lončarić (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Tomislav Petković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-08, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U ovome radu uspoređivane su dvije različite arhitekture neuronskih mreža za segmentaciju biomedicinskih slika. Slike iz korištenog skupa podataka su dobivene pomoću X-PCI metode snimanja srca štakora. Arhitekture mreža koje se koriste su U-net i Tiramisu. Segmentacija slike je proces dijeljenja digitalne slike na više smislenih, logično izabranih skupova piksela ili segmenata. Obje korištene mreže su konvolucijske mreže. Prema dobivenim rezultatima zaključuje se da točnost segmentacije U-net mreže manje ovisi o dimenzijama slika od Tiramisu. Tiramisu mreža ima puno dulje trajanje treniranja, te na nju puno više utječu dimenzije slika. Implementirana rješenja tog problema su prepolavljanje dimenzija slike i dijeljenje slike na četiri slike jednakih dimenzija. Mreže su uspoređivane pomoću Dice i IoU. Najbolji dobiveni rezultati su od Tiramisu mreže trenirane na skupu podataka slika podijeljenih na četiri jednakih dimenzija (Dice = 0.9540, IoU = 0.9112).

Abstract (english)

This work compares the effectiveness of two different neural network architectures for biomedical image segmentation. Images from the data set used were obtained using the X-PCI method on the heart of a rat. The architectures used in this paper are U-net and Tiramisu. Image segmentation is the process of dividing a digital image into multiple meaningful, logically chosen sets of pixels or segments. Both networks used are convolutional networks. According to the obtained results, it is concluded that the accuracy of U-net network segmentation depends less on the image dimensions than Tiramisu. The Tiramisu network has a much longer training duration, and it is much more affected by the dimensions of the images. Implemented solutions to this problem are halving the image dimensions and dividing the image into four images of equal dimensions. The best results were obtained from Tiramisu networks trained on a set of images divided into four equal dimensions (Dice = 0.9540, IoU = 0.9112).

• Keywords: Segmentacija; biomedicina; neuronska mreža; konvolucija; U-net; Tiramisu; Dice; IoU
• Keywords (english): Segmentation; biomedicine; neural network; convolution; U-net; Tiramisu; Dice; IoU
• Language: croatian
• URN:NBN: urn:nbn:hr:168:540585
• 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 10:25:35