Electronic Theses and Dissertation

Segmentasi citra MRI jantung merupakan komponen penting dalam pengembangan sistem analisis medis berbasis otomasi untuk diagnosis dan evaluasi penyakit kardiovaskular. Penelitian ini mengusulkan model segmentasi citra MRI jantung yang mengintegrasikan arsitektur U-Net dengan Jaringan Adverserial Generatif (GAN) untuk meningkatkan ketepatan batas anatomi dan konsistensi spasial hasil segmentasi. Eksperimen dilakukan menggunakan dataset Automated Cardiac Diagnosis Challenge (ACDC) 2017, yang terdiri dari citra MRI jantung beserta anotasi ground truth dari ahli sebagai acuan evaluasi. Tahapan penelitian meliputi preprocessing data, pembangunan model, pelatihan, dan evaluasi kuantitatif. Kinerja segmentasi diuji menggunakan metrik Dice Score (DSC) dan Intersection over Union (IoU). Model baseline U-Net dengan konfigurasi batch size 4, learning rate 1×10⁻⁴, optimizer Adam, dan fungsi loss Categorical Cross-Entropy menghasilkan rata-rata DSC sebesar 71,99% dan IoU sebesar 62,73%. Model U-Net+GAN yang diusulkan menunjukkan peningkatan kinerja secara signifikan dengan rata-rata DSC sebesar 80,50% dan IoU sebesar 72,46%. Hasil ini menunjukkan bahwa penggabungan mekanisme adversarial mampu meningkatkan realisme anatomi pada mask hasil prediksi dan secara substansial memperbaiki akurasi segmentasi dibandingkan model U-Net baseline.

Cardiac MRI image segmentation is a crucial component in the development of automated medical analysis systems for the diagnosis and evaluation of cardiovascular diseases. This study proposes a cardiac MRI image segmentation model that integrates the U-Net architecture with a Generative Adversarial Network (GAN) to improve anatomical boundary accuracy and spatial consistency of the segmentation results. Experiments were conducted using the Automated Cardiac Diagnosis Challenge (ACDC) 2017 dataset, which consists of cardiac MRI images along with expert-annotated ground truth masks for evaluation. The research stages include data preprocessing, model development, training, and quantitative evaluation. Segmentation performance was assessed using the Dice Score (DSC) and Intersection over Union (IoU) metrics. The baseline U-Net model with a batch size of 4, a learning rate of 1×10⁻⁴, the Adam optimizer, and Categorical Cross-Entropy loss achieved an average DSC of 71.99% and an IoU of 62.73%. The proposed U-Net+GAN model demonstrated a significant performance gain, achieving an average DSC of 80.50% and IoU of 72.46%. These results indicate that the incorporation of adversarial learning enhances anatomical realism in the predicted masks and substantially improves segmentation accuracy compared to the baseline U-Net model.