A fatty liver diseases classification network based on adaptive coordination attention with label smoothing

Abstract

Accurate classification of fatty liver disease (FLD) from fuzzy boundary areas and lesion areas in ultrasound imaging is crucial for clinical diagnosis and treatment. This paper presents an efficient convolutional neural network based on adaptive coordinated attention (ACA), which combines label smoothing and EfficientNet for FLD classification. We only uses pixels and labels of original brightness mode ultrasound images as inputs, and achieves four classification of FLD. It avoids the decoupling of feature extraction and classification, and achieves end-to-end automatic classification identification. Based on ACA, the model can adaptively encode lateral and vertical positional information adaptively into channel attention that adaptively learns information in both spatial directions to enhance the representation of the objects of concern. Furthermore, the interpretability of the model is demonstrated through feature heatmap and hybrid heatmap visualization techniques. The experimental results reveal that the proposed approach attains an accuracy of 94.96% and an AUC of 0.97 on the private dataset. The F1 scores for the four categories are 0.963, 0.942, 0.932, and 0.955, respectively, contributing to an overall F1 score of 0.955, surpassing other baseline models and achieving a new state-of-the-art (SOTA). Meanwhile, this approach also achieves SOTA accuracy of 97.82% on public datasets. The proposed approach has surpassed the manual recognition of fatty liver brightness mode ultrasound images in terms of its cognitive ability, which helps to aid the expert diagnosis. It has better performance and achieves end-to-end classification compared to other methods.