CLAUNet-LFB0-CBAM a deep learning framework for explainable liver fibrosis stage classification from ultrasound images

Abstract

Automated liver fibrosis staging through ultrasound images is still difficult owing to low contrast, speckle artifacts, and minor differences between successive stages. We propose a hybrid deep learning architecture named CLAUNet-LFB0-CBAM for multi-class liver fibrosis classification from ultrasound images. The method uses a preprocessing pipeline step followed by an augmented EfficientNet-B0 network with CBAM. Moreover, to deal with class imbalance issue, the method employs weighted cross-entropy loss and weighted random sampling, combined with augmentation techniques. The proposed network is compared with baseline architectures like VGG-16, ResNet-50, DenseNet-121, and EfficientNet-B0 within identical experimental conditions. Experimental results show that the proposed method provides superior results in terms of 99.05% accuracy, 0.980 F1-score, and 0.9987 AUC, surpassing all evaluated baseline architectures under identical experimental conditions. Importantly, the performance improvement is achieved mainly in intermediate liver fibrosis stages which are crucial for clinical evaluation. Grad-CAM analysis supports that the network learns the appropriate features from anatomical liver regions. Moreover, the network achieves lightweight implementation with only around 4.8 million parameters.