Skripsi

GENERATIVE AI BERBASIS DEEP LEARNING PADA ANALISIS CITRA ULTRASONOGRAFI

---

Ultrasound (USG) image analysis for detecting Congenital Heart Disease (CHD), such as Atrial Septal Defect (ASD), Ventricular Septal Defect (VSD), and Atrioventricular Septal Defect (AVSD), is still limited by the scarcity of clinical datasets. This study evaluates three Generative Adversarial Network (GAN) architectures, namely Deep Convolutional GAN (DCGAN), Wasserstein GAN with Gradient Penalty (WGAN-GP), and StyleGAN2-ADA, to generate synthetic data for augmentation purposes. Based on the Frechet Inception Distance (FID), Kernel Inception Distance (KID), and Learned Perceptual Image Patch Similarity (LPIPS) metrics, StyleGAN2-ADA achieved the best image quality. An instance segmentation YOLOv11 model was then optimized through hyperparameter tuning using real data to obtain the optimal configuration, which was consistently applied across three training scenarios: real data, synthetic data, and their combinations (1:1 to 4:1). The evaluation results show that real data achieved the best unseen performance with a Mask mAP50 of 0,5068, while all GAN-based scenarios failed to surpass the baseline under out-of-distribution conditions. These findings indicate that although the synthetic images have high visual quality, the per-patient synthetic data distribution is still insufficient to represent the anatomical variability of new patients, highlighting the need for advanced approaches such as cross-patient GAN training to improve clinical robustness.

---

Availability

Detail Information

Series Title

-

Call Number

T2009922026

Publisher

Indralaya : Prodi Sistem Komputer, Fakultas Ilmu Komputer Universitas Sriwijaya., 2026

Collation

xiv, 103 hlm.; ilus.; tab.; 29 cm.

Language

Indonesia

ISBN/ISSN

-

Classification

006.310 7

Content Type

Text

Media Type

-

Carrier Type

-

Edition

-

Subject(s)

Prodi Sistem Komputer
Analisis Data Komputer--Deep Learning

Specific Detail Info

-

Statement of Responsibility

MI

Other version/related

File Attachment

Comments

---

You must be logged in to post a comment