Klasifikasi dan Segmentasi Tingkat Keganasan Tumor Otak Glioma dengan Modalitas Magnetic Resonance Imaging (MRI) Berbasis Deep Learning

Irena, Florencia (2025) Klasifikasi dan Segmentasi Tingkat Keganasan Tumor Otak Glioma dengan Modalitas Magnetic Resonance Imaging (MRI) Berbasis Deep Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5023211009-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (12MB) | Request a copy

Abstract

Berdasarkan Global Cancer Statistic 2022, terdapat sejumlah 321.731 kasus tumor otak di seluruh dunia, dengan jenis kasus paling umum yaitu glioma yang mendominasi sebanyak 30% dari total keseluruhan. World Health Organization (WHO) mengelompokkan tingkat keganasan tumor glioma menjadi Low Grade Gliomas atau LGG (grade 1 dan 2) dan High Grade Gliomas atau HGG (grade 3 dan 4). Penentuan grade glioma sangat penting untuk memprediksi tingkat kelangsungan hidup pasien. Namun, analisis berbasis MRI cukup kompleks dan memakan waktu. Oleh karena itu, dibutuhkan sistem Computer-Aided Diagnosis (CAD) untuk mendukung pengambilan keputusan klinis. Metode CAD yang paling umum digunakan adalah machine learning (ML). Namun metode ini sangat bergantung pada pemilihan fitur secara manual yang dapat memengaruhi ketahanan model. Deep learning (DL) menjadi alternatif yang mampu melakukan ekstraksi fitur serta segmentasi tumor secara otomatis. Penelitian ini mengusulkan multi-task learning yang menggabungkan proses segmentasi dan ekstraksi fitur untuk klasifikasi glioma grade 2, 3, dan 4 menggunakan citra MRI multi-sequence. Model segmentasi DenseNet121-UNet dikembangkan dengan strategi fine-tuning dan penambahan dropout untuk meningkatkan generalisasi. Model ini menunjukkan performa stabil tanpa overfitting, dengan nilai DSC sebesar 0,9673 dan loss 0,0364 pada data uji. Kemudian fitur deep diekstraksi dari beberapa lapisan encoder menggunakan GAP dan GMP, serta dilakukan ekstraksi fitur radiomics. Lalu dilakukan pengujian model SVM dengan berbagai kombinasi fitur dan jenis kernel. Penggunaan fitur deep dengan kernel polynomial menghasilkan performa terbaik. Parameter optimal (C = 0,1; gamma = 0,01; coef0 = 1; degree = 3) berhasil mencapai akurasi sebesar 90,21% hanya dalam waktu 10,2 detik. Hasil ini menunjukkan efektivitas integrasi segmentasi dan klasifikasi dalam satu alur kerja untuk prediksi tingkat keganasan glioma.
================================================================================================================================
According to the Global Cancer Statistics 2022, there were approximately 321,731 cases of brain tumors worldwide, with gliomas being the most common type, accounting for 30% of the total cases. The World Health Organization (WHO) classifies gliomas based on their malignancy grades into LGG (grades 1 and 2) and HGG (grades 3 and 4). Glioma grading is vital for predicting survival rates. However, MRI-based analysis is complex and time-consuming. So, there is a need for an Computer-Aided Diagnosis (CAD) system to support clinical decision making, yet Machine Learning (ML) methods that rely on manually selected features can affect model robustness. In addition, Deep Learning (DL) can automatically extract features and segment tumors. This study proposes a multi task learning framework that integrates segmentation and feature extraction for classifying glioma grades 2, 3, and 4 using multi-sequence MRI images. A DenseNet121-UNet segmentation model was developed with fine-tuning strategies and the incorporation of dropout layers to enhance generalization. The model demonstrated stable performance without overfitting, achieving a DSC of 0.9673 and a loss of 0.0364. Deep features were extracted from multiple encoder layers using Global Average Pooling and Global Max Pooling. Then, Support Vector Machine (SVM) classifier was evaluated using various feature combinations and kernel types. The usage of deep features with a Polynomial kernel yielded the best performance. The optimal model (C = 0.1; gamma = 0.01, coef0 = 1, degree = 3) achieved an accuracy of 90.21% within just 10.2 seconds. These findings demonstrate the effectiveness of integrating segmentation and classification into a unified workflow for glioma malignancy grade prediction.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deep Learning, DenseNet121, Glioma, SVM, U-Net Deep Learning, DenseNet121, Glioma, SVM, U-Net
Subjects:	T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models. T Technology > T Technology (General) > T58.62 Decision support systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Biomedical Engineering > 11410-(S1) Undergraduate Thesis
Depositing User:	Florencia Irena
Date Deposited:	04 Aug 2025 02:54
Last Modified:	04 Aug 2025 02:54
URI:	http://repository.its.ac.id/id/eprint/126154

Actions (login required)

View Item