Evaluasi Normalisasi pada Model 3D U-NetR untuk Segmentasi Aneurisma Otak

Jauhari, Ahmad Alvin (2026) Evaluasi Normalisasi pada Model 3D U-NetR untuk Segmentasi Aneurisma Otak. Project Report. [s.n.], [s.l.]. (Unpublished)

Text 5025221180-Project_Report.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Kerja praktik ini bertujuan melakukan analisis fine-tuning terhadap jenis normalisasi pada sistem segmentasi citra medis yang digunakan dalam platform BrainNav. Proses fine-tuning dilakukan menggunakan model UNETR (U-Net with Transformers) dari library MONAI (Medical Open Network for AI) dengan membandingkan tiga jenis normalisasi, yaitu Batch Normalization, Instance Normalization, dan Group Normalization. Evaluasi performa model dilakukan menggunakan beberapa metrik segmentasi, meliputi Total Loss, Dice Similarity Value, Surface Dice Value, Hausdorff Distance, False Negative Rate, dan Volumetric Similarity. Hasil pengujian menunjukkan bahwa Batch Normalization unggul pada empat metrik, yaitu Total Loss dengan nilai 0,352625, Dice Similarity Value dengan nilai 0,3434, Surface Dice Value dengan nilai 0,438215, dan Predicted Positive Voxel sebanyak 3347721 pada epoch terbaiknya. Sementara itu, Instance Normalization memberikan performa terbaik pada dua metrik, yaitu Hausdorff Distance dan Volumetric Similarity, dengan nilai Haussdorff Distance sebesar 118,0775 dan Volumetric Similarity sebesar 0,54899 pada epoch terbaiknya. Selain itu, pada epoch terbaik berdasarkan metrik Dice Similarity Value, model dengan Batch Normalization menghasilkan jumlah prediksi benar tertinggi dibandingkan dua jenis normalisasi lainnya. Berdasarkan hasil evaluasi kuantitatif tersebut, model hasil fine-tuning dengan konfigurasi normalisasi terbaik dipilih untuk diimplementasikan pada platform BrainNav guna mendukung pengembangan sistem segmentasi citra otak yang lebih optimal.
====================================================================================================================================
This internship aims to conduct a fine-tuning analysis of the normalization types used in the medical image segmentation system used in the BrainNav platform. The fine-tuning process was performed using the UNETR (U-Net with Transformers) model from the MONAI (Medical Open Network for AI) library, comparing three types of normalization: Batch Normalization, Instance Normalization, and Group Normalization. Model performance was evaluated using several segmentation metrics, including Total Loss, Dice Similarity Value, Surface Dice Value, Hausdorff Distance, False Negative Rate, and Volumetric Similarity. Test results showed that Batch Normalization excelled in four metrics: Total Loss with a value of 0.352625, Dice Similarity Value with a value of 0.3434, Surface Dice Value with a value of 0.438215, and Predicted Positive Voxels with 3347721 in its best epoch. Meanwhile, Instance Normalization performed best in two metrics: Hausdorff Distance and Volumetric Similarity, with a Hausdorff Distance value of 118.0775 and a Volumetric Similarity value of 0.54899 in its best epoch. Furthermore, in the best epoch based on the Dice Similarity Value metric, the Batch Normalization model produced the highest number of correct predictions compared to the other two normalization types. Based on the quantitative evaluation results, the fine-tuned model with the best normalization configuration was selected for implementation on the BrainNav platform to support the development of a more optimal brain image segmentation system.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Brain Nav, Fine tuning, Machine learning, Normalisasi, UNETR
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Ahmad Alvin Jauhari
Date Deposited:	14 Jan 2026 04:07
Last Modified:	14 Jan 2026 04:07
URI:	http://repository.its.ac.id/id/eprint/129590

Actions (login required)

View Item