SicklePrenatal: Biometric-Guided Deep Learning For Early Sickle Cell Risk Detection From Fetal Ultrasound

Bisiriyu Maqsudhat Kofoworola

Keywords: Low-Resource Healthcare, Ultrasound, Deep Learning, Attention Mechanisms, Medical Imaging

Abstract:

Sickle Cell Disease (SCD) is one of the most prevalent inherited disorders in Sub-Saharan Africa, yet limited prenatal screening leads to high infant mortality. We present SicklePrenatal, an attention-based deep learning framework for early, non-invasive SCD risk prediction from fetal ultrasound in low-resource settings.

Our approach integrates two complementary streams: (1) a structure-preserving encoder enhanced with Sobel-Laplacian filters for vascular edge detection, and (2) a biometric encoder informed by gestational age and key fetal measurements (head circumference, femur length, abdominal diameter). A cross-attention module fuses both branches, capturing clinically relevant patterns under weak supervision.

To address the lack of labeled datasets, we design a proxy-based annotation pipeline grounded in clinical heuristics and literature-backed thresholds. We further pretrain with a contrastive self-supervised objective on grayscale ultrasound frames, and fine-tune on a dataset of 3,382 scans from a regional health clinic.

SicklePrenatal achieves 76.2% accuracy, surpassing baseline CNNs (63.7%) and ResNet classifiers (69.1%), while preserving interpretability through saliency maps localized around high-risk vascular zones. The model is optimized for edge deployment, enabling inference on low-compute devices.

Our work contributes a scalable, low-cost framework for prenatal SCD screening, with potential to generalize to other hemoglobinopathies and fetal-risk disorders, advancing maternal-fetal healthcare in underserved populations.

Submission Number: 246