Semantic segmentation of the spine based on magnetic resonance imaging

U sklopu ovog rada obrađena je tema računalnog vida za medicinsku svrhu. U medicini se koriste različite metode za segmentiranje dijelova tijela sa snimaka (npr. organi, kosti, razna tkiva..) i ovaj rad fokusira se na segmentaciju kralježnice snimaka magnetske rezonance pomoću algoritama strojnog vida. To znači da je program sam sposoban prepoznati svaki kralježak zasebno, te ih označiti. Ovim putem se cijeli postupak tumačenja sadržaja MRI snimaka može znatno olakšati i ubrzati. Postoje metode koje imaju tradicionalan pristup segmentacije, te one koje rade pomoću dubokih neuronskih mreža. Istraženo je više metoda u oba pristupa segmentaciji, kao što su CNN, FCN, U-Net, GAN, Edge detection, thresholding,.. Kasnije u radu segmentirane su dvije različite kralježnice otprilike istim pristupom, obje pomoću algoritama obrade slike. Dobiveni rezultati jedne od tih dviju segmentacija, uspoređene su pomoću Dice koeficijenta sličnosti sa State-of-the-art rješenjima segmentacije. Evaluirana je uspješnost segmentacije, te su doneseni zaključci. Cilj ovog rada je istražiti koliko se dobro može provesti postupak segmentacije pristupom obrade slike u usporedbi s dobro razvijenim već postojećim metodama segmentacije u medicini.This paper focuses on the topic of computer vision for medical purposes. In medicine, various methods are employed for the segmentation of body parts from images, such as organs, bones, and various tissues. This paper specifically concentrates on segmenting the spine from magnetic resonance imaging (MRI) scans using machine vision algorithms. This means that the program is capable of autonomously recognizing and labeling each vertebra individually. This approach significantly streamlines and accelerates the process of interpreting the content of MRI scans. There are methods that follow a traditional segmentation approach, as well as those that utilize deep neural networks. Several methods have been explored in both segmentation approaches, including CNN, FCN, U-Net, GAN, edge detection, thresholding, and others. Later in the paper, two different spines were segmented using roughly the same image processing algorithms. The obtained results of one of these two segmentations were compared using the Dice coefficient of similarity with state-of-the-art segmentation solutions. The success of the segmentation was evaluated, and conclusions were drawn. The aim of this paper is to investigate how well the segmentation process can be conducted using image processing approaches compared to well-established existing methods of segmentation in medicine