Segmentacija stanica na mikroskopskim slikama zasnovana na modelima dubokog učenja

master's thesis

• Title (english): Cell segmentation on microscopic images based on deep learning models
• Author: Hrvoje Ćošković
• Mentor: Tomislav Šmuc (mentor)
• Committee member: Tomislav Šmuc (predsjednik povjerenstva)
• Committee member: Zvonimir Bujanović (član povjerenstva)
• Committee member: Mladen Jurak (član povjerenstva)
• Committee member: Goran Igaly (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2021-03-04, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

U uvodu ovoga diplomskog rada mogu se pronaći valjani razlozi okretanja strojnom i dubokome učenju, njegove sve veće prednosti nad tradicionalnim metodama, ali i pokoja mana. Zatim smo teoretski skupa prošli cijeli proces dubokog učenja, od modeliranja problema, samog procesa učenja pa i do najsitnijih gradivnih jedinica svake neuronske mreže, njenih slojeva i principa rada i učenja. Nakon toga smo se okrenuli konkretnoj primjeni dubokog učenja u medicini, segmentiranju slika stanica nastalih svjetlosnom elektromikroskopijom stanica iz različitih vrsta tkiva koja su se našla na natjecanju Data Science Bowlu 2018. Vidjeli smo i opisali najuspješnije algoritme i metode koje su natjecatelji koristili, a slobodno možemo reći i najsuvremenija rješenja koja su neki od njih implementirali. Zatim smo došli i do najvažnijeg dijela ovoga rada, procesa modeliranja autorovog modela koji se okušao u segmentaciji slika tumora na centralnom živčanom sustavu kućnog miša. Zadnje poglavlje obuhvaća sve korake i procese u nastanku istoga, te rezultate koji su na nekim uzorcima na jako dobrom tragu, ali kod pojedinih ipak sugeriraju da bi se trebali okrenuti kompleksnijim modelima većeg kapaciteta učenja.

Abstract (english)

In the introduction of this Master work, one can find valid reasons for turning to machine and deep learning, its increasing advantages over traditional methods, but also a few defects. Then, theoretically, we went through the entire process of deep learning, from problem modeling, the learning process itself to the tiniest building units of each neural network, its layers and principles of work and learning. After that, we turned to the concrete application of deep learning in medicine, segmentation of images of cells produced by light electromicroscopy of cells from different types of tissues found in the Data Science Bowl competition in 2018. We have seen and described the most successful algorithms and methods used by competitors, and we can freely say state-of-the-art solutions that some of them have implemented. Then we came to the most important part of this paper, the modelling process of the author’s model that was tested in segmentation of tumor images on the central nervous system of the mouse. The last chapter covers all the steps and processes in the creation of the same, as well as the results that are on some samples on a very good track, but in some cases suggest that they should turn to more complex models of greater learning capacity.

• Keywords: strojno i duboko učenje; neuronske mreže; svjetlosna elektromikroskopija stanica
• Keywords (english): deep learning; neural network; light electromicroscopy of cells
• Language: croatian
• URN:NBN: urn:nbn:hr:217:645201
• Study programme: Title: Computer Science and Mathematics; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra računarstva i matematike (magistar/magistra računarstva i matematike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2021-03-24 11:25:22