Gornale, Shivanand, Kamat, Priyanka, Hiremath, Prakash and Siddalingappa, Rashmi ORCID: https://orcid.org/0000-0001-9786-8436 (2025) A Hybrid Ensemble of Denoising Autoencoders and Deep Learning Models for Fetal Image Analysis. Cureus Journal Of Computer Science.

Preview

Text Cureus_Sep2025_Fetus_Anomaly.pdf - Published Version Available under License Creative Commons Attribution. Download (1MB) | Preview

Abstract

Medical image analysis, particularly ultrasonography, has involved increasing attention in computer science and engineering due to its potential for automated and scalable interpretation. Ultrasound imaging is widely used in prenatal care because of its non-invasive nature and cost-effectiveness. Automated analysis of fetal ultrasound images can improve diagnostic accuracy and reduce inter-observer variability. However, challenges such as speckle noise, low contrast, and anatomical variations across trimesters make automated interpretation difficult, requiring robust preprocessing, segmentation, and classification methods.

This study proposes a hybrid ensemble deep learning framework for analyzing fetal ultrasound images. The framework integrates a denoising autoencoder for noise reduction and image enhancement, as well as seven segmentation architectures (U-Net, DeepLabV3+, DenseNet-U-Net, MFP-UNet, Attention U-Net, MobileNet-U-Net, and ResNet-U-Net), and five ensemble strategies (maximum voting, majority voting, weighted voting, confidence-based fusion, and averaging) to enhance segmentation performance. A multi-input classification approach is also introduced, combining individual and ensemble segmentation outputs in a fine-tuned DenseNet121 for trimester categorization (first, second, and third trimesters) based on head circumference and femur length.

The framework is evaluated using Dice score, mean intersection over union, accuracy, precision, recall, and F1-score. Experimental results show that ensemble strategies significantly improve segmentation. The multi-input classification achieves 92.50% accuracy for head circumference and 90.60% for femur length on the custom dataset, as well as 83.68% on the HC18 dataset, outperforming individual models.

The main contributions include (1) a hybrid ensemble strategy for robust segmentation and (2) a multi-input trimester classification method. The proposed framework is generalizable and can be extended to other medical imaging applications beyond fetal ultrasound analysis.

Item Type:	Article
Status:	Published
DOI:	10.7759/s44389-025-09506-x
Subjects:	Q Science > Q Science (General) T Technology > T Technology (General)
School/Department:	London Campus
URI:	https://ray.yorksj.ac.uk/id/eprint/12901

University Staff: Request a correction | RaY Editors: Update this record