MUCM-FLLs: Multimodal ultrasound-based classification model for focal liver lesions

Abstract

Timely detection and accurate classification of focal liver lesions (FLLs) are crucial for improving patient survival rates and providing optimal treatment strategies. This study proposes a multimodal ultrasound-based classification model (MUCM-FLLs) to assist clinicians in efficiently leveraging multimodal ultrasound data for FLL diagnosis. We utilized data from 359 patients with histopathologically confirmed FLLs to develop a model that integrates lesion B-mode ultrasound images, background liver ultrasound images, color Doppler flow imaging, and clinical data. Incremental modality experiments were conducted, demonstrating average classification accuracies of 55.0%, 54.2%, 61.8%, and 83.7% for single-mode to four-mode configurations. These results highlight the effectiveness of combining multiple modalities and reveal differing sensitivities of various diseases to specific modalities. Cross-validation further validated the model’s robustness and generalizability, confirming the advantages of multimodal diagnosis. During training, we introduced a gradient adjustment strategy with a learning score metric to address learning rate disparities among modalities under multimodal data training. This strategy effectively mitigated imbalances in modality optimization, ensuring that each modality received adequate training. Additionally, we quantitatively analyzed the contributions of different modalities to the diagnosis of various diseases and calculated inter-modality weights, significantly improving the model’s predictive accuracy. Supported by these strategies, MUCM-FLLs achieved an overall accuracy of 92.2%. This study highlights the potential of multimodal fusion and optimization strategies to enhance the diagnostic performance of FLLs and provides significant technical support for clinical diagnosis.