Classification of Ultrasound Images Using ResNet-50 with a Convolutional Block Attention Module (CBAM)

Abstract

Liver fibrosis staging is a crucial component in the clinical management of chronic liver disease because it directly affects prognosis, therapeutic decision-making, and long-term patient monitoring. Ultrasound imaging is widely used as a noninvasive diagnostic modality due to its safety, low cost, and broad accessibility. Nevertheless, ultrasound-based fibrosis assessment remains challenging because liver parenchymal echotexture often exhibits low contrast, speckle noise, and subtle inter-stage variations, particularly among adjacent METAVIR stages. These characteristics frequently limit the effectiveness of conventional convolutional neural networks, which tend to emphasize dominant global patterns while suppressing weak but clinically meaningful texture cues. This study presents a task-oriented integration of a Convolutional Block Attention Module into a ResNet-50 backbone to enhance feature discrimination for five-stage liver fibrosis classification using heterogeneous B-mode ultrasound images. Rather than introducing a new attention mechanism, the contribution lies in the systematic insertion of CBAM after residual outputs across multiple network stages, enabling repeated channel and spatial recalibration from low-level texture descriptors to higher-level semantic representations. To further improve robustness and reduce prediction variance, a stratified 5-fold training strategy is combined with logit-level ensemble inference, where logits from independently trained fold models are averaged prior to Softmax normalization. Experiments were conducted on a publicly available dataset comprising 6,323 ultrasound images acquired from two tertiary hospitals using multiple ultrasound systems, with fibrosis stages labeled from F0 to F4 according to histopathology-based METAVIR scoring. The proposed framework achieves a test accuracy of 98.34%and consistently high precision, recall, and F1 scores across all fibrosis stages, with the most pronounced improvement observed for intermediate stages. Statistical analysis based on paired fold-wise comparisons confirms that the performance gain over the baseline ResNet 50 model is statistically significant. These results demonstrate that combining lightweight attention-based feature refinement with logit ensemble inference effectively addresses the inherent challenges of ultrasound-based liver fibrosis staging and provides a reliable noninvasive decision support framework with strong potential for clinical application and future multicenter validation.