Physics-Informed Neural Networks for Ultrasound-Based Varicose Vein Screening

Abstract

Early screening of varicose veins using non-invasive and low-cost sensing techniques remains a practical challenge in community and home-based healthcare settings. Conventional ultrasound imaging systems are accurate but require bulky hardware, skilled operators, and clinical environments, which limit accessibility for large-scale screening. This study proposes a physics-informed deep learning framework for one-dimensional ultrasound A-scan analysis to enable lightweight and interpretable varicose vein screening. A convolutional neural network with an encoder–decoder reconstruction branch and a classification head was developed. The reconstruction branch incorporates physics-motivated constraints, including second-order temporal smoothness and signal energy consistency, to regularize the learned waveform representation. These constraints aim to preserve physically meaningful propagation characteristics while suppressing noise and overfitting. A stratified train–test split with a guaranteed test size of 50 samples was employed to improve statistical reliability under small-sample conditions. Experimental results on 75 labeled A-scan segments demonstrated stable convergence and high screening performance. The final model achieved approximately 98% test accuracy with zero false negatives, indicating strong sensitivity for detecting abnormal vascular conditions. The physics-informed reconstruction produced smoother yet structurally consistent waveforms compared with raw signals, suggesting improved interpretability of learned representations. The findings indicate that physics-guided learning can enhance robustness and clinical relevance in small-sample ultrasound screening tasks. While further validation with participant-level separation and larger cohorts is required, the proposed framework provides a feasible direction for portable, AI-assisted vascular screening in non-clinical environments.

Keywords: Physics-informed Neural Networks, Ultrasound Screening, Varicose Veins, Health Informatics, Biomedical Signal Analysis

DOI: 10.54941/ahfe1007495