Automated Maternal–Fetal Health Analysis through Deep Neural Network Integration in Ultrasound Imaging

Gornale, Shivanand S.; Kamat, Priyanka; Siddalingappa, Rashmi; Li, Kefang; Wen Goh, Khang

Automated Maternal–Fetal Health Analysis through Deep Neural Network Integration in Ultrasound Imaging

Gornale, Shivanand S., Kamat, Priyanka, Siddalingappa, Rashmi ORCID: https://orcid.org/0000-0001-9786-8436, Li, Kefang and Wen Goh, Khang (2026) Automated Maternal–Fetal Health Analysis through Deep Neural Network Integration in Ultrasound Imaging. International Journal of Computer Information Systems and Industrial Management Applications, 18 (5s). pp. 1038-1068.

Preview

Text
84(2852).pdf - Published Version
Available under License Creative Commons Attribution.
| Preview

Official URL: https://doi.org/10.70917/ijcisim-2026-2852

Abstract

Ultrasonography is a widely used, non-invasive modality for monitoring fetal development and identifying growth abnormalities during pregnancy. This study proposes an attention-enhanced deep learning framework for automated fetal ultrasound image classification, addressing two tasks: (i) biometric parameter-based classification into four categories Head Circumference (HC), Femur Length (FL), Abdominal Circumference (AC), and Crown–Rump Length (CRL); and (ii) trimester-based classification into three stages using HC, FL, and the publicly available HC18 benchmark dataset. The framework employs a feature-level ensemble fusion strategy by integrating 1024-dimensional feature representations from four pretrained convolutional neural networks, ResNet50, VGG16, VGG19, and Xception, producing a unified 4096-dimensional fused vector to capture complementary and hierarchical features. Gradient-weighted Class Activation Mapping (Grad-CAM) is applied as an attention-enhanced preprocessing step using a strictly frozen ImageNet-pretrained VGG16, generating spatially weighted heatmap overlays that are superimposed onto the original fetal ultrasound images and used as enriched RGB inputs to the ensemble classifier, improving both classification performance and model interpretability. The entire pipeline, including Grad-CAM generation, dataset partitioning, data augmentation, and ensemble classification, operates automatically with no manual intervention at any stage. Data augmentation is applied exclusively to the training partition after dataset splitting, ensuring no data leakage into the validation or test sets. Experimental results demonstrate that the proposed fused ensemble model achieves 99.63% accuracy for biometric parameter-based classification, and 97.21%, 98.58%, and 92.72% for HC-based, FL-based, and HC18-based trimester classification, respectively, outperforming all individual baseline models. These findings confirm the effectiveness of combining Grad-CAM attention-enhanced inputs with deep ensemble feature fusion for robust, interpretable, and generalizable fetal ultrasound analysis in clinical decision support systems.