Abstract
Recently, contrast-enhanced ultrasound (CEUS) has presented a potential value in the diagnosis of liver trauma, the leading cause of death in blunt abdominal trauma. However, the inherent speckle noise and the complicated visual characteristics of blunt liver trauma in CEUS images make the diagnosis highly dependent on the expertise of radiologists, which is subjective and time-consuming. Moreover, the intra- and inter-observer variance inevitably influences the accuracy of diagnosis using CEUS. In this study, we propose a Label-Noisy-Resistant CNN-Transformer Hybrid Architecture (LNRHA) for CUES liver trauma classification. Firstly, a CNN-Transformer-based Self-Contextual Dual Transformer (SCDT) module, a shared feature encoder followed by the dual-perspective Transformer-based modules, is developed to perceive the semantics of trauma lesions from neighbor-contextual and self-attention perspectives. Moreover, to mitigate the annotation noise due to intra- and inter-observer variance, we design a Confidence-Based Label Filter (CLF) module to distinguish potential label noise data based on the ensemble of the SCDT. The uncertainty of the detected noisy data is gradually penalized using a newly designed loss function, making full use of all the data while avoiding overfitting to misleading information, thus improving the classification performance. Extensive experimental results on an in-house liver trauma CEUS dataset show that our network architecture can achieve promising performance. Significantly, the experimental results of our LNRHA method on label noise data also outperform most state-of-the-art classification methods, suggesting its effectiveness in diagnosing liver trauma.