Akselerasi Deep Learning Untuk Segmentasi Volumetrik Citra Medis Dengan Multi-GPU Distributed Data Parallel

Kurniawan, Richo Yudha (2026) Akselerasi Deep Learning Untuk Segmentasi Volumetrik Citra Medis Dengan Multi-GPU Distributed Data Parallel. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221242-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Beban komputasi konvolusi 3D menjadikan pelatihan model segmentasi volumetrik citra medis pada konfigurasi single-GPU memerlukan waktu yang lama, sementara karakteristik scaling implementasi Distributed Data Parallel (DDP) pada beban kerja ini belum dikarakterisasi secara sistematis dalam literatur. Tugas Akhir ini menginvestigasi secara empiris speedup, scaling efficiency, dan trade-off antara akselerasi waktu dan akurasi segmentasi yang diukur menggunakan Dice Similarity Coefficient (DSC) dan 95% Hausdorff Distance (HD95) pada implementasi DDP. Arsitekur 3D U-Net dilatih pada dataset Brain Tumor Segmentation (BraTS) 2024 Adult Glioma Post-treatment, dengan pendekatan strong scaling pada konfigurasi 1, 2, dan 4 GPU NVIDIA A40. Hasil Tugas Akhir ini menunjukkan speedup sebesar 1,9697× pada konfigurasi 2 GPU dan 3,8614× pada konfigurasi 4 GPU terhadap baseline 1 GPU, dengan scaling efficiency masing-masing sebesar 98,48% dan 96,53%. Selisih DSC antarkonfigurasi GPU pada keenam wilayah tumor tidak melebihi 0,0199, jauh di bawah standar deviasi per kasus sebesar ±0,36, dan uji Wilcoxon signed-rank tidak menemukan perbedaan yang bermakna secara praktis di seluruh 36 pasangan konfigurasi GPU yang diuji pada masing-masing metrik evaluasi. Hasil ini menunjukkan bahwa implementasi DDP pada segmentasi volumetrik 3D dapat mencapai akselerasi waktu pelatihan yang mendekati linear tanpa mengalami degradasi akurasi yang berarti.
============================================================================================================================================
The computational burden of 3D convolution makes training volumetric medical image segmentation models on single-GPU configurations prohibitively time-intensive, yet the scaling characteristics of Distributed Data Parallel (DDP) implementations under this workload remain insufficiently characterized in the literature. This study empirically investigates the speedup, scaling efficiency, and trade-off between training acceleration and segmentation accuracy, measured via Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (HD95), of a DDP implementation for 3D volumetric segmentation. A 3D U-Net was trained on the Brain Tumor Segmentation (BraTS) 2024 Adult Glioma Post-Treatment dataset under a strong scaling regime across 1, 2, and 4 NVIDIA A40 GPU configurations. Training speedups of 1.9697× and 3.8614× were achieved at 2 and 4 GPUs relative to the single-GPU baseline, with scaling efficiencies of 98.48% and 96.53%, respectively. DSC differences across GPU configurations on all six tumor subregions did not exceed 0.0199, a value well within the per-case standard deviation of ±0.36, and the Wilcoxon signed-rank test found no practically meaningful differences on either evaluation metric across all 36 GPU configuration pairs tested. These results demonstrate that DDP-based training of 3D volumetric segmentation models can achieve near-linear acceleration without meaningful accuracy degradation.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	3D U-Net, BraTS 2024, Distributed Data Parallel, Segmentasi Volumetrik Citra Medis, Strong Scaling, 3D U-Net, BraTS 2024, Distributed Data Parallel, Strong Scaling, Volumetric Medical Image Segmentation
Subjects:	T Technology > T Technology (General) > T385 Visualization--Technique T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T58.62 Decision support systems T Technology > T Technology (General) > T58.64 Information resources management T Technology > T Technology (General) > T58.8 Productivity. Efficiency T Technology > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering
Depositing User:	Richo Yudha Kurniawan
Date Deposited:	28 Jul 2026 06:44
Last Modified:	28 Jul 2026 06:44
URI:	http://repository.its.ac.id/id/eprint/138663

Actions (login required)

View Item