International Research Journal of Engineering and Technology (IRJET) Volume: 12 Issue: 11 | Nov 2025
www.irjet.net
e-ISSN: 2395-0056 p-ISSN: 2395-0072
Automated Detection and Segmentation of Fetal Brain Abnormalities Using Real-Time YOLO Deep Learning Architecture Renuka Devi1, Ritesh2, Veeresh3, Ajay Walikar4, Praveen Kyade5 1Professor, Department of CSE, Government Engineering College, Bidar, Karnataka, India 2345Student, B.E, Government Engineering College, Bidar, Karnataka, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Early identification of structural anomalies in
operator-dependent variability and limits consistency. Recent advances in deep learning, particularly real-time object detection methods like YOLO, offer an opportunity to automate and standardize diagnostic workflows. This study develops an optimized YOLO-based framework designed for rapid and accurate detection of fetal brain abnormalities in prenatal ultrasound images.
fetal brain tissue via prenatal ultrasonography constitutes a fundamental prerequisite for appropriate perinatal intervention strategies and informed parental decisionmaking. This investigation introduces a specialized deep learning framework leveraging the YOLO (You Only Look Once) real-time object detection paradigm, engineered to autonomously identify and spatially localize multiple categories of fetal central nervous system abnormalities in Bmode ultrasound imagery. The proposed methodology incorporates domain-specific image preprocessing techniques, hierarchical multi-scale feature representation, and customized loss function formulations optimized for medical imaging contexts. Comprehensive evaluation across a clinically-sourced dataset comprising 2,847 annotated prenatal ultrasound studies demonstrates detection accuracy of 96.5%, precision of 95.8%, recall of 94.7%, and mean average precision of 91.8%, significantly surpassing established baseline approaches including Faster R-CNN and Mask R-CNN. The framework successfully identifies diverse pathological entities including ventricular enlargement (ventriculomegaly), agenesis of the corpus callosum, cerebellar developmental deficiency, open neural tube pathology, and choroid plexus cystic formations. The sub100ms inference latency enables practical deployment as a real-time clinical decision-support modality, facilitating standardized, operator-independent assessment protocols during routine prenatal screening. This contribution advances the intersection of deep learning methodologies and obstetric diagnostic imaging, establishing a reproducible, scalable, and clinically-translatable framework for enhancing detection sensitivity and reducing inter-observer variability in fetal neurosonography.
1.1 Research Motivation and Contributions Clinical motivation arises from limitations of current diagnostic practices, including prolonged examination time, inter-observer variation, and risks of diagnostic oversight. The primary contributions of this research include: specialized YOLO optimization tailored for medical ultrasound preprocessing, extensive comparative evaluation against Faster R-CNN, Mask R-CNN, and traditional machinelearning baselines, rigorous validation using a dataset of 2,847 annotated clinical ultrasound examinations, and development of a clinically oriented integration framework for obstetric diagnostic workflows.
1.2 Technological Evolution and Motivation Recent advancements in convolutional neural network (CNN) architectures and single-stage object detection methodologies have catalyzed a paradigm shift in medical image analysis. The YOLO framework, first introduced by Redmon et al. and subsequently refined through multiple iterations (YOLOv3-v8), processes images in a unified, endto-end manner rather than generating intermediate region proposals, thereby achieving substantial computational efficiency gains while maintaining or improving detection accuracy metrics.
Key Words: Fetal brain abnormality detection, YOLO, deep learning, medical image segmentation, computeraided diagnosis, prenatal diagnostics
Previous applications of deep learning to fetal imaging have predominantly employed two-stage detectors or standard CNN classification pipelines, which, while achieving reasonable accuracy levels (typically 85-92%), exhibit computational overhead unsuitable for real-time clinical workflows or require extensive post-processing. The emergence of optimized YOLO variants addresses these constraints through integrated architectural innovations including feature pyramid networks, spatial pyramid pooling, and enhanced loss functions. This investigation hypothesizes that thoughtfully adapted YOLO architecture, coupled with preprocessing optimizations specific to
1.INTRODUCTION Prenatal diagnosis of fetal brain structural anomalies is a crucial element of obstetric imaging. Early detection of abnormalities such as ventriculomegaly, corpus callosum agenesis, cerebellar hypoplasia, and neural tube defects supports appropriate maternal counseling and perinatal management. Conventional ultrasound-based diagnosis depends heavily on expert interpretation, which introduces
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 872