Segmentasi Citra MRI Tumor Glioblastoma Berbasis Modifikasi U-Net Untuk Lokalisasi Tumor Praoperasi

Akib, Muhammad Qusay Yubasyrendra (2025) Segmentasi Citra MRI Tumor Glioblastoma Berbasis Modifikasi U-Net Untuk Lokalisasi Tumor Praoperasi. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5023211039-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB)

Abstract

Glioblastoma multiforme (GBM) adalah tumor otak primer paling agresif dengan struktur morfologi yang kompleks dan batas spasial yang buram. Proses segmentasi citra MRI secara manual masih umum digunakan dalam praktik klinis untuk keperluan diagnosis dan perencanaan pra-operasi, namun sangat bergantung pada keahlian radiolog serta rentan terhadap bias subjektif dan kesalahan, terutama pada area kecil seperti enhancing tumor (ET). Oleh karena itu, diperlukan pendekatan segmentasi otomatis yang akurat dan efisien untuk mengatasi kompleksitas segmentasi tumor GBM. Penelitian ini mengusulkan pengembangan sistem segmentasi otomatis berbasis modifikasi U-Net 2D yang mengintegrasikan Multi-Scale Residual Block (MSRB) di encoder untuk menangkap fitur spasial multi-skala, serta Channel Attention dan External Attention untuk memperkuat seleksi fitur semantik dan konteks spasial global. Sebagai penyempurna, tahap postprocessing berbasis ET Boundary Refinement diterapkan untuk meningkatkan ketepatan prediksi batas pada subregion ET. Dataset BraTS 2021 digunakan sebagai dasar pelatihan yang selanjutnya diolah melalui tahapan preprocessing citra meliputi centroid-based cropping, multi-sequence fusion (FLAIR, T1, T1CE), normalisasi intensitas, dan peningkatan kontras menggunakan CLAHE untuk menghasilkan input yang konsisten dan informatif. Evaluasi performa dilakukan menggunakan metrik Dice Score dan Hausdorff Distance95 (HD95). Hasil eksperimen menunjukkan bahwa model akhir dengan kombinasi Multi-Scale Residual Block (MSRB), attention mechanism, dan ET Boundary Refinement mampu mencapai peningkatan performa yang signifikan pada ketiga subregion tumor. Diperoleh nilai Dice Score 0.98 untuk whole tumor (WT), 0.98 untuk tumor core (TC), dan 0.91 untuk enhancing tumor (ET), dengan nilai Hausdorff Distance 95 (HD95) masing-masing sebesar 1.0 untuk WT, 1.0 untuk TC, dan 2.2 untuk ET.
==================================================================================================================================
Glioblastoma multiforme (GBM) is the most aggressive primary brain tumor, characterized by complex morphological structures and blurred spatial boundaries. Although manual MRI image segmentation is commonly used in clinical practice for diagnostic and preoperative planning purposes, it is highly dependent on the expertise of the radiologist and is prone to subjective bias and errors, particularly in small areas such as enhancing tumors (ET). Therefore, an accurate, efficient, automatic segmentation approach is needed to address the complexity of GBM tumor segmentation. This study proposes developing an automatic segmentation system based on a modified 2D U-Net. This system integrates Multi-Scale Residual Blocks (MSRBs) in the encoder to capture multi-scale spatial features. It also uses Channel Attention and External Attention to enhance semantic feature selection and global spatial context. An ET boundary refinement-based postprocessing stage is also applied to improve the accuracy of boundary predictions in ET subregions. The BraTS 2021 dataset was used for training and was processed through preprocessing stages, including centroid-based cropping, multi-sequence fusion (FLAIR, T1, and T1CE), intensity normalization, and contrast enhancement using CLAHE, to produce consistent and informative inputs. Performance was evaluated using Dice score and Hausdorff distance 95 (HD95) metrics. The results of the experiment show that the final model, which combines a Multi-Scale Residual Block (MSRB), an attention mechanism, and ET Boundary Refinement, achieved significant performance improvements in all three tumor subregions. The Dice scores obtained were 0.98 for the whole tumor (WT), 0.98 for the tumor core (TC), and 0.91 for the enhancing tumor (ET), with HD95 values of 1.0 for WT and TC and 2.2 for ET.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Glioblastoma multiforme, segmentasi otomatis, U-Net 2D, Multi-Scale Residual Block, attention mechanism, ET Boundary Refinement Glioblastoma multiforme, automatic segmentation, 2D U-Net, Multi-Scale Residual Block, attention mechanism, ET Boundary Refinement
Subjects:	R Medicine > R Medicine (General) > R858 Deep Learning R Medicine > RC Internal medicine > RC78.7.N83 Magnetic resonance imaging. T Technology > T Technology (General) > T385 Visualization--Technique T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5102.9 Signal processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Biomedical Engineering > 11410-(S1) Undergraduate Thesis
Depositing User:	Muhammad Qusay Yubasyrendra Akib
Date Deposited:	04 Aug 2025 02:27
Last Modified:	04 Aug 2025 02:27
URI:	http://repository.its.ac.id/id/eprint/125639

Actions (login required)

View Item