Segmentasi 3D Data Anatomi Otak Menggunakan Metode Transformator U-Net Berbasis 3D Point Cloud Berdasarkan Metode GAN

Muzada Elfa, Muhammad Ridho (2024) Segmentasi 3D Data Anatomi Otak Menggunakan Metode Transformator U-Net Berbasis 3D Point Cloud Berdasarkan Metode GAN. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025221025-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (3MB) | Request a copy

Abstract

Penelitian ini merupakan pendekatan medis terbaru yang bertujuan untuk rekonstruksi 3D dan segmentasi bentuk anatomi otak corpus callosum, bertujuan untuk membantu mengangkat batasan visualisasi untuk prosedur dengan lingkungan visual 2D yang terbatas. Data yang digunakan pada penelitian ini merupakan citra MRI dari National Hospital dan Rumah Sakit dr. Soetomo di Surabaya berjumlah 45 data. Penelitian ini memiliki dua tahapan, pertama yaitu rekonstruksi 2D slice Magnetic Resonance Imaging (MRI) otak menjadi 3D MRI dengan metode Generative Adversarial Network (GAN) yang setelahnya akan di segmentasi dengan metode trasnformator U-Net untuk mendapatkan hasil segmentasi yang akurat tanpa adanya noise.
Hasil dari generated slice dengan metode GAN dibandingkan dengan metode trilinear interpolasi, yang mana kedua hasil tersebut diuji untuk menghitung keakuratan hasil generated slice dibandingkan dengan ground truth. Metrik evaluasi yang digunakan untuk pengujian yaitu Structured Similarity Index Measure (SSIM) dan Peak-Signal-Nation-Ratio (PSNR). Hasil yang di dapatkan yaitu SSIM 0,978 dan PSNR 68db untuk metode GAN lalu SSIM 0,964 dan PSNR 66db untuk metode trilinear tnterpolasi, hasil visualisasi dari metode GAN unggul dibandingkan metode trilinear interpolasi.
Tahapan selanjutnya yaitu segmentasi 3D MRI dengan metode yang digunakan yaitu U-Net. Hasil segmentasi dibandingkan dengan ground truth melalui pengujian metrik evaluasi dice coefficient F1 Score, SSIM dan Mean Squared Error (MSE). Hasil yang ada dibandingkan dari percobaan empat pengujian dengan input hyperparameter berbeda yaitu, (jumlah epoch, dropout, batch normalization dan filtersize), nilai tertinggi didapat pada pengujian ke empat dengan hasil akurasi F1-Score 92,16 persen, SSIM 0,995 dan MSE Error 0,005, dengan jumlah epoch 50, dropout rate 0,5 dan filtersize 64. Hasil Segmentasi divisualisasikan dengan open3D menggunakan hasil numpy array 2D MRI tiap slice yang ditumpuk menjadi 3D MRI menghasilkan 3D Point Cloud dari corpus callosum.

==============================================================

=============================

This research is a novel medical approach aimed at 3D reconstruction and segmentation of the anatomical shape of the brain's corpus callosum, aiming to help lift visualization limitations for procedures with a limited 2D visual environment. The data used in this research are MRI images from the National Hospital and Dr. Soetomo in Surabaya totaled 45 data. This research has two stages, the first is the reconstruction of 2D Magnetic Resonance Imaging (MRI) slices of the brain into 3D MRI using the Generative Adversarial Network (GAN) method, which will then be segmented using the U-Net transformer method to get accurate segmentation results without any noise.
The results from generated slices using the GAN method are compared with the trilinear interpolation method, where both results are tested to calculate the accuracy of the generated slice results compared to ground truth. The evaluation metrics used for testing are Structured Similarity Index Measure (SSIM) and Peak-Signal-Nation-Ratio (PSNR). The results obtained were SSIM 0.978 and PSNR 68db for the GAN method, then SSIM 0.964 and PSNR 66db for the trilinear interpolation method, the visualization results from the GAN method were superior to the trilinear interpolation method.
The next stage is 3D MRI segmentation with the method used, namely U-Net. The segmentation results are compared with ground truth through testing the dice coefficient F1 Score, SSIM and Mean Squared Error (MSE) evaluation metrics. The existing results were compared from four test experiments with different hyperparameter input, namely, (number of epochs, dropout, batch normalization and filter size), the highest value was obtained in the fourth test with F1-Score accuracy results of 92.16 percent, SSIM 0.995 and MSE Error 0.005 , with a number of epochs of 50, a dropout rate of 0.5 and a filter size of 64. The segmentation results were visualized with open3D using the results of a 2D MRI numpy array for each slice which was stacked into a 3D MRI to produce a 3D Point Cloud of the corpus callosum.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Deep Learning, Magnetic Resonance Images, Otak, 3D Point Cloud, Rekonstruksi, U-Net, Segmentasi, =========================================================== Brain, Deep Learning, Magnetic Resonance Images, 3D Point cloud, Reconstruction, U-Net, Segmentation
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Muhammad Ridho Muzada Elfa
Date Deposited:	07 Aug 2024 08:17
Last Modified:	07 Aug 2024 08:17
URI:	http://repository.its.ac.id/id/eprint/114007

Actions (login required)

View Item