Arsitektur Dense-UNet dengan Transfer Learning untuk Segmentasi Dinding Ventrikel Kiri pada Citra Ultrasound Jantung

Unggul, Didik Bani (2024) Arsitektur Dense-UNet dengan Transfer Learning untuk Segmentasi Dinding Ventrikel Kiri pada Citra Ultrasound Jantung. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003212005-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 April 2026. Download (8MB) | Request a copy

Abstract

Dinding ventrikel kiri sering menjadi region of interest (ROI) dalam pemeriksaan ultrasound jantung. Hal ini dikarenakan abnormalitas bagian tersebut dapat menjadi indikator dini berbagai penyakit kardiovaskular. Meskipun demikian, citra ultrasound dengan resolusi yang rendah menjadikan pengamatan dinding ventrikel kiri sulit dilakukan. Masalah ini mendorong pengembangan metode segmentasi otomatis yang mampu menandai area dinding ventrikel kiri dengan lebih jelas, akurat, dan cepat daripada pengamatan manual. Penelitian ini mengajukan pendekatan deep learning untuk menjawab tantangan tersebut. Secara lebih detail, penelitian ini mengkonstruksi tiga arsitektur Dense-UNet hasil perpaduan U-Net dan Dense-Net yang dinamai Dense-UNet-121, Dense-UNet-169, dan Dense-UNet-201. Penambahan tujuh skenario transfer learning (NoTL, TLS1, TLS2, TLS3_20F, TLS3 _40F, TLS3_60F, dan TLS3_80F) juga dilakukan untuk menangani kebutuhan data training yang besar. Kombinasi tiga varian Dense-UNet dan tujuh skenario tersebut dibandingkan dengan arsitektur state-of-the-art, yaitu U-Net, pada dataset publik dari Hamad Medical Corporation, Qatar University, dan Tampere University (HMC QU). Hasil evaluasi dengan Intersection-over-Union (IoU) menunjukan bahwa 95% model Dense-UNet lebih bagus dari model U-Net. Penggunaan transfer learning juga diketahui dapat meningkatkan IoU dan mempercepat durasi training. Skenario TLS3 secara konsisten menjadi yang terbaik saat training dengan lonjakan kinerja teramati pada saat melewati cutoff freeze-unfreeze. Skenario tersebut juga mendominasi jajaran model dengan IoU validasi tertinggi di masing-masing arsitektur. Model Dense-UNet-201 dengan skenario TLS3_20F ditetapkan sebagai model terbaik karena memiliki IoU validasi tertinggi sebesar 0,6227. Hasil visualisasi di data testing menegaskan kemampuan model dalam memprediksi bentuk dan posisi ROI yang mirip dengan ground truth. Model tersebut lalu diintegrasikan ke dalam suatu Graphical User Interface (GUI) yang diujicobakan pada data testing. Proses tersebut memberikan rata-rata durasi segmentasi sebesar 30,4 detik per video, mengindikasikan GUI dapat melakukan segmentasi otomatis dengan cepat.
==============================================================================================================================
The left ventricular wall often becomes a region of interest (ROI) in cardiac ultrasound examinations. This is because abnormalities in this area can serve as early indicators of various cardiovascular diseases. However, low-resolution ultrasound images make it challenging to observe the left ventricular wall. This issue has driven the development of automated segmentation methods capable of annotating the left ventricular wall area more clearly, accurately, and quickly than manual observation. This research proposes a deep learning approach to address this challenge. In more detail, this study constructs three Dense-UNet architectures from the hybridization of U-Net and Dense-Net, named Dense-UNet-121, Dense-UNet-169, and Dense-UNet-201. Additionally, seven transfer learning scenarios (NoTL, TLS1, TLS2, TLS3_20F, TLS3_40F, TLS3_60F, and TLS3_80F) are implemented to handle the large training data requirements. The combinations of the three Dense-UNet variants and the seven scenarios are compared with the state-of-the-art architecture, i.e. U-Net, on public datasets from Hamad Medical Corporation, Qatar University, and Tampere University (HMC QU). Evaluation results using Intersection-over-Union (IoU) indicate that 95% of Dense-UNet models outperform the U-Net model. The use of transfer learning is also found to enhance IoU and expedite training duration. The TLS3 scenario consistently performs the best during training, with a noticeable performance spike after the cutoff freeze-unfreeze phase. This scenario also dominates the top model with the highest validation IoU in each architecture. The Dense-UNet-201 model with the TLS3_20F scenario is identified as the best model, achieving the highest validation IoU of 0.6227. Visualization results on testing data confirm the model's ability to predict ROI shapes and positions similar to the ground truth. The model is then integrated into a Graphical User Interface (GUI) and tested on testing data. This process yields an average segmentation duration of 30.4 seconds per video, indicating that the GUI can perform automatic segmentation quickly.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Dense-UNet, Dinding Ventrikel Kiri, Segmentasi, Transfer Learning, Left Ventricular Wall, Segmentation
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA76.9.U83 Graphical user interfaces. User interfaces (Computer systems)--Design. R Medicine > R Medicine (General) > R858 Deep Learning R Medicine > RC Internal medicine > RC78.7.U4 Ultrasonic imaging.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49101-(S2) Master Thesis
Depositing User:	Didik Bani Unggul
Date Deposited:	09 Feb 2024 00:07
Last Modified:	09 Feb 2024 00:07
URI:	http://repository.its.ac.id/id/eprint/106449

Actions (login required)

View Item