SEGMENTASI INSTANCE MULTICLASS UNTUK INTERPRETASI OBJEK CITRA JANTUNG JANIN

SYAPUTRA, HADI and Nurmaini, Siti and Partan, Radiyati Umi (2025) SEGMENTASI INSTANCE MULTICLASS UNTUK INTERPRETASI OBJEK CITRA JANTUNG JANIN. Doctoral thesis, Sriwijaya University.

Preview	Image RAMA_21001_03043681823004_cover.jpg - Accepted Version Available under License Creative Commons Public Domain Dedication. Download (221kB) | Preview
	Text RAMA_21001_03043681823004.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (3MB) | Request a copy
	Text RAMA_21001_03043681823004_TURNITIN.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (1MB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_01_front_ref.pdf - Accepted Version Available under License Creative Commons Public Domain Dedication. Download (865kB)
	Text RAMA_21001_03043681823004_0002085908_0017077206_02.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (443kB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_03.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (993kB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_04.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (968kB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_05.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (143kB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_ref.pdf - Bibliography Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (265kB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_lamp.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (4MB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_06_ref.pdf - Bibliography Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (265kB) | Request a copy
	Text RAMA_21001_03043681823004_0002085908_0017077206_07_lamp.pdf - Accepted Version Restricted to Repository staff only Available under License Creative Commons Public Domain Dedication. Download (4MB) | Request a copy

Abstract

This study aims to develop a multiclass instance segmentation model based on Mask RCNN to identify anatomical structures of the normal fetal heart in four-chamber view (4CV) ultrasound images. The background of this research stems from the urgent need for early detection of congenital heart disease (CHD), one of the leading causes of neonatal mortality, combined with the limited availability of fetomaternal subspecialists and the variability in ultrasound image quality encountered in clinical practice. The proposed model is designed as a clinical decision support system powered by artificial intelligence (AI) to assist obstetricians in the initial interpretation of fetal cardiac ultrasound images, particularly in healthcare facilities with limited specialist resources. The dataset used in this study comprises 176 images, extracted from fetal echocardiography videos and annotated into 10 anatomical classes, including major structures (LV, LA, RV, RA, AO), minor structures (PV, MV, TV, AV), and an additional class for the spine, which serves as a reference for fetal anatomical orientation. The study also investigates the impact of various image enhancement techniques, including Histogram Equalization (HE), Adaptive Histogram Equalization (AHE), Contrast Limited AHE (CLAHE), Blind Deblurring (BD), and their combinations, on the segmentation performance. Evaluation was conducted using standard performance metrics, namely mean Average Precision (mAP), Intersection over Union (IoU), Dice Similarity Coefficient (DCS), and Confusion Matrix. Experimental results show that the R50_sgd_20 model configuration achieved the best segmentation performance across most major anatomical classes. However, challenges remain in segmenting small anatomical structures such as Pulmonary Valve (PV) and Aortic Valve (AV). Although enhancement techniques improved image clarity, their effect on segmentation accuracy varied across classes. These findings suggest that segmentation performance is influenced not only by image quality, but also by the underlying model architecture, class distribution, and spatial characteristics of the anatomical structures within the image.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Instance Segmentation, Mask R-CNN, Fetal Heart, Ultrasound, Deep Learning, Image Enhancement (CLAHE, HE, BD), mAP, IoU, DCS, Confusion Matrix
Subjects:	Q Science > Q Science (General) > Q334-342 Computer science. Artificial intelligence. Algorithms. Robotics. Automation.
Divisions:	03-Faculty of Engineering > 21001-Engineering Science (S3)
Depositing User:	Hadi Syaputra
Date Deposited:	14 Oct 2025 03:19
Last Modified:	14 Oct 2025 03:22
URI:	http://repository.unsri.ac.id/id/eprint/182190

Actions (login required)

View Item