Segmentasi Glioma Otak Pada Citra MRI Menggunakan Multi-Modal Selective Kernel 3D U-Net Dengan Imbalance-Aware Tversky Loss

Lathifa, Arsya Dewi (2026) Segmentasi Glioma Otak Pada Citra MRI Menggunakan Multi-Modal Selective Kernel 3D U-Net Dengan Imbalance-Aware Tversky Loss. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221015-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (26MB) | Request a copy

Abstract

Segmentasi glioma otak pada citra MRI multi-modal menghadapi tantangan ketidakseimbangan kelas yang ekstrem, terutama pada region tumor berukuran kecil seperti necrotic tumor core (NCR). NCR hanya menempati 0,04% dari total volume citra. Selain itu, arsitektur segmentasi 3D standar masih memiliki keterbatasan dalam mengadaptasi receptive field terhadap variasi morfologi sub-region serta mengintegrasikan informasi spesifik antar-modalitas. Penelitian ini mengusulkan arsitektur Multi-Modal Selective Kernel 3D U-Net (MMSK-3D U-Net). Arsitektur usulan merupakan model 3D U-Net yang dilengkapi blok Selective Kernel berbasis cross-modal gating untuk adaptasi receptive field, dan menggunakan Tversky loss untuk mengatasi ketidakseimbangan kelas. Evaluasi dilakukan pada dataset BraTS 2024 GLI melalui studi ablasi lima skenario untuk mengisolasi kontribusi setiap komponen. Model Usulan memiliki rata-rata Dice Similarity Coefficient (DSC) per-kelas sebesar 0,802 dan rata-rata DSC sub-region sebesar 0,851. Pada tingkat sub-region, DSC yang diperoleh adalah 0,902/0,873/0,779 untuk WT/TC/ET. Analisis kontribusi komponen menunjukkan bahwa modifikasi arsitektur melalui mekanisme cross-modal gating pada MMSK meningkatkan sensitivitas rata-rata dari 0,823 menjadi 0,836 dibandingkan 3D U-Net standar dengan fungsi loss yang sama. Meskipun demikian, dari studi ablasi diketahui bahwa faktor paling dominan dalam pemulihan performa kelas minoritas adalah optimasi fungsi loss. Penerapan Tversky loss berhasil memulihkan pengenalan kelas NCR dari yang sebelumnya gagal dipelajari oleh Dice loss dari DSC 0,001 menjadi DSC 0,705. Konfigurasi parameter alpha=0,3 dan beta=0,7 yang diadaptasi dari literatur menunjukkan performa yang lebih baik dibandingkan konfigurasi dengan penalti false negative yang lebih berat (alpha=0,2, beta=0,8), sehingga konfigurasi tersebut dipertahankan pada penelitian ini. Penelitian ini menyimpulkan bahwa pada kasus ketidakseimbangan kelas ekstrem, optimasi fungsi loss jauh lebih menentukan keberhasilan segmentasi, sementara arsitektur MMSK memberikan kontribusi dalam meningkatkan sensitivitas dan menjaga kestabilan pembelajaran pada kelas minoritas.
=====================================================================================================================================
Glioma segmentation on multi-modal brain MRI faces the challenge of extreme class imbalance, particularly for small tumor regions such as the necrotic tumor core (NCR). NCR occupies only 0.04% of the total image volume. In addition, standard 3D segmentation architectures remain limited in adapting their receptive field to the morphological variation of tumor sub-regions and in integrating modality-specific information. This study proposes the Multi-Modal Selective Kernel 3D U-Net (MMSK-3D U-Net) architecture. The proposed architecture is a 3D U-Net equipped with a Selective Kernel block based on cross-modal gating for receptive field adaptation, and employs a Tversky loss to address class imbalance. Evaluation was conducted on the BraTS 2024 GLI dataset through an ablation study of five scenarios to isolate the contribution of each component. The proposed model achieved a mean per-class Dice Similarity Coefficient (DSC) of 0.802 and a mean sub-region DSC of 0.851. At the sub-region level, the obtained DSC values were 0.902/0.873/0.779 for WT/TC/ET. Analysis of component contributions shows that the architectural modification through the cross-modal gating mechanism in MMSK improved the mean sensitivity from 0.823 to 0.836 compared with a standard 3D U-Net using the same loss function. Nevertheless, the ablation study reveals that the most dominant factor in recovering minority-class performance is the optimization of the loss function. The application of Tversky loss successfully recovered the recognition of the NCR class, which had previously failed to be learned under Dice loss, from a DSC of 0.001 to a DSC of 0.705. The parameter configuration alpha=0.3 and beta=0.7, adapted from the literature, showed better performance than a configuration with a heavier false negative penalty (alpha=0.2, beta=0.8), and was therefore retained in this study. This study concludes that under extreme class imbalance, the optimization of the loss function is far more decisive for segmentation success, while the MMSK architecture contributes to improving sensitivity and maintaining learning stability for the minority class.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Segmentasi Glioma, MMSK-3D U-Net, Selective Kernel, Cross-Modal Gating, Tversky Loss, Ketidakseimbangan Kelas. Glioma Segmentation, MMSK-3D U-Net, Selective Kernel, Cross-Modal Gating, Tversky Loss, Class Imbalance
Subjects:	R Medicine > R Medicine (General) > R858 Deep Learning R Medicine > RC Internal medicine > RC0254 Neoplasms. Tumors. Oncology (including Cancer) R Medicine > RC Internal medicine > RC78.7.N83 Magnetic resonance imaging. T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Arsya Dewi Lathifa
Date Deposited:	24 Jul 2026 07:09
Last Modified:	24 Jul 2026 07:09
URI:	http://repository.its.ac.id/id/eprint/137145

Actions (login required)

View Item