Metode za proširenje skupa slika srca

undergraduate thesis

• Title (english): Data Augmentation Methods for Cardiac Images
• Author: Bruno Jerković
• Mentor: Sven Lončarić (mentor)
• Committee member: Sven Lončarić (predsjednik povjerenstva)
• Committee member: Jelena Ivanković (član povjerenstva)
• Committee member: Karlo Koščević (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-06, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U ovom radu razmotrile su se dvije metode za proširenje skupa slika za treniranje neuronske mreže te se ispitalo kako njihova implementacija utječe na jednom primjeru neuronske mreže koja obavlja semantičku segmentaciju slika presjeka srca.
Prva metoda podrazumijeva korištenje osnovnih transformacija nad slikama kao što su: zrcaljenje, rotacija i translacija.
Druga metoda koristi implementaciju neuronske mreže generativnog suparničkog modela (GAN) koji će generirati nove slikovne podatke. GAN... More je model neuronske mreže koji se sastoji od dvije neuronske mreže koje se međusobno „natječu“. Prva, generator, pokušava generirati slike što sličnije onima u originalnom skupu podataka, a druga, diskriminator, pokušava razaznati slike stvorene od strane generatora od onih iz početnog podatkovnog skupa.
U radu su spomenute i metrike za evaluaciju neuronskih mreža koje obavljaju prepoznavanje uzoraka na slikama; IoU i Dice. Oboje su izvedeni kao jednadžbe površina koje je mreža dobro klasificirala i površina koje je mreža pogrešno klasificirala.
Pri implementaciji metoda koristile su se razne tehnologije, kao na primjer Pytorch, biblioteka sa funkcijama potrebnima za izradu neuronskih mreža. GAN je implementiran promjenama raznih hiper-parametara i slojeva dok se nije postigao željeni rezultat.
Na kraju su se metode implementirale na jednom primjeru neuronske mreže. Prema rezultatima je zaključeno da na tom primjeru samostalna implementacija metoda nije dovoljna te da je pokazano da samostalna implementacija metoda bez mijenjanja hiper-parametara i slojeva u arhitekturi mreže može negativno utjecat na njezin rad.
Less

Abstract (english)

In this paper, two methods for expanding a set of images were considered and it was examined how their implementation affects one example of a neural network that performs semantic segmentation of cross-sectional images of the heart.
The first method involves the use of basic transformations in images such as: mirroring, rotation and translation.
The second method uses the implementation of a neural network of a generative adversarial model (GAN) that generates new images. GAN is a neural... More network model that consists of two neural networks that "compete" against each other. The first one, the generator, tries to generate images as similar as possible to those in the original data set, and the second one, the discriminator, tries to distinguish the images created by the generator from those in the initial data set.
The paper also mentions metrics for evaluating neural networks that perform pattern recognition in images; IoU and Dice. Both are derived as equations of areas that are well classified by the network and areas that are incorrectly classified by the network.
Various technologies were used to implement the methods, such as Pytorch, a library with the functions needed to build neural networks. GAN was implemented by changing various hyper-parameters and layers until the desired result was achieved.
Finally, the methods were implemented on one example of a neural network. According to the results, it was concluded that in this example the independent implementation of methods is not enough and that it is shown that independent implementation of methods without changing hyper-parameters and layers in the network architecture can negatively affect its operation.
Less

• Keywords: umjetna inteligencija; podatkovni skup; GAN; neuronska mreža; IoU; Dice
• Keywords (english): artificial intelligence; dataset; GAN; neural network; IoU; Dice
• Language: croatian
• URN:NBN: urn:nbn:hr:168:655402
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 15:10:25