LDCE-Net: A Custom-Built Lightweight CNN for Classifying Liver Fibrosis Stages via Ultrasound Imaging

Abstract

Liver fibrosis is a progressive disease that, undetected, proceeds to cirrhosis and liver failure. Early detection through ultrasound imaging remains a challenging task with subtle grayscale shifts and reliance on experienced interpretation. Here, we present LDCE-Net, a lightweight tailored convolutional neural network architecture developed ab initio to discriminate liver ultrasound images into three relevant clinical classes: Normal, Fibrosis, and Cirrhosis. Unlike previous approaches relying on pretrained models, LDCE-Net was meticulously developed and trained to extract both shallow and deep features with a double-path structure amalgamated with a feature attention module. Testing was conducted on a labeled dataset with cross-validation and achieved a test accuracy of 87%, with balanced F1-scores on each class. Learning curves and validation curves both confirmed steady learning with minimal overfitting, and the confusion matrix highlighted class-wise high performance with specific ability to distinguish healthy and heavily fibrotic tissue. The trained model was also incorporated within a user-centric Streamlit web interface with real-time prediction of fibrosis stage with ultrasound input. Such findings warrant LDCE-Net’s potential for a practical, accurate, and accessible screening device against liver disease.

Keywords: Liver Fibrosis, Ultrasound Imaging, Lightweight CNN, Deep Learning.