Widodo, Ramadhan Sanyoto Sugiharso (2024) Sintesis Citra dengan Paired dan Unpaired GAN untuk Segmentasi Citra MRI Jantung Berbasis Transfer Learning. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6022231035-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6022231035-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (4MB) | Request a copy |
Abstract
Penyakit kardiovaskular merupakan salah satu penyebab utama kematian di dunia, termasuk di Indonesia. Pemeriksaan atrium kiri jantung yang tepat menggunakan pencitraan medis sangat penting untuk prognosis, diagnosis, dan perawatan penyakit kardiovaskular. Namun, keterbatasan dataset citra medis menjadi tantangan dalam pelatihan model segmentasi otomatis menggunakan deep learning. Generative Adversarial Network (GAN) telah menunjukkan potensi dalam pencitraan medis, tetapi penerapan paired dan unpaired GAN serta transfer learning masih belum dieksplorasi secara mendalam. Penelitian ini bertujuan untuk mengatasi keterbatasan dataset citra medis dengan mensintesis citra MRI jantung beranotasi atrium kiri menggunakan paired dan unpaired GAN sebagai data tambahan. Tujuan utamanya adalah untuk mengembangkan model segmentasi yang efektif dan mengurangi ketergan tungan pada anotasi manual, serta meningkatkan kemampuan generalisasi model melalui transfer learning. Hasil penelitian sintesis GAN dengan metrik evaluasi Structural Similarity Index Measure (SSIM) dan Peak Signal to Noise Ratio (PSNR) menunjukkan bahwa paired GAN (Pix2Pix) menghasilkan label sintesis yang sangat mirip dengan citra asli dengan SSIM 0,978 dan PSNR 29,712, sedangkan unpaired GAN (CycleGAN) lebih fleksibel untuk data terbatas dalam menghasilkan sintesis citra dengan SSIM 0,693 dan PSNR 23,899. Transfer learning terbukti meningkatkan stabilitas pelatihan dan kemampuan generalisasi model. Kombinasi data tambahan dan transfer learning meningkatkan performa model segmentasi dengan rata-rata dice score, IoU score, dan sensitivitas masing-masing di atas 0,95, 0,94, dan 0,96 pada mayoritas skenario model segmentasi. Penggunaan paired dan unpaired GAN untuk sintesis data citra medis, dikombinasikan dengan transfer learning, memberikan solusi efektif untuk mengatasi keterbatasan dataset dan meningkatkan performa model segmentasi.
=======================================================================================================================================
Cardiovascular disease is one of the leading causes of death in the world, including in Indonesia. Proper examination of the left atrium of the heart using medical imaging is essential for the prognosis, diagnosis, and treatment of cardiovascular diseases. However, the limitation of medical image datasets is a challenge in training automated segmentation models using deep learning. Generative Adversarial Networks (GANs) have shown potential in medical imaging, but the application of paired and unpaired GANs and transfer learning has not been explored in depth. This study aims to overcome the limitations of medical image datasets by synthesizing left atrial annotated cardiac MRI images using paired and unpaired GANs as additional data. The main goal is to develop an effective segmentation model that reduces the dependency on manual annotation and improves the model’s generalization ability through transfer learning. The results of GAN synthesis research with the evaluation metrics Structural Similarity Index Measure (SSIM) and Peak Signal to Noise Ratio (PSNR) show that paired GAN (Pix2Pix) produces synthesized labels that are very similar to the original image with SSIM 0.978 and PSNR 29.712. At the same time, unpaired GAN (CycleGAN) is more f lexible for limited data in producing synthesized images with SSIM 0.693 and PSNR 23.899. Transfer learning was shown to improve training stability and model generalization ability. The combination of additional data and transfer learning improved the segmentation model performance with average Dice Score, IoU Score, and sensitivity above 0.95, 0.94, and 0.96, respectively, in most segmentation model scenarios. Using paired and unpaired GANs for medical image data synthesis, combined with transfer learning, provides an effective solution to overcome dataset limitations and improve segmentation model performance.
Actions (login required)
 |
View Item |