BC-DCNN: A Novel Method to Denoise CEUS Images Combining Bidirectional ConvLSTM with 3D DnCNN

Background and objectives: Ultrasound diagnosis technology is widely used in medical diagnosis due to its advantages of low cost, timeliness, and no radiation. However, the scattered noise in the image hinders the doctor's accurate diagnosis. Besides, the current research on ultrasonic video denoising is still in the preliminary stage, and it is difficult to effectively denoise while fully retaining detail feature.Therefore,the objective of this study is to propose a novel denosing framework capable of generating clearer images. Methods: In this paper, we propose a novel framework named BC-DCNN (Bidirectional ConvLSTM with 3D DnCNN), which combines bidirectional ConvLSTM (Convolutional Long Short-Term Memory) with 3D DnCNN is proposed to generate clearer images. First, before entering the network, the CEUS video will pass through an abnormal frame detection module to filter out abnormal frames, and combine the synthetic ultrasound video as a training dataset, thereby preventing the model from overfitting and improving the generalization ability of the network. Second, we introduce a 3D convolution operator to extend the 2D DnCNN denoising model to 3D space to process image sequences. In addition, the model learns the temporal features of the video through the bidirectional ConvLSTM and takes full advantage of the inter-frame information. At the same time, temporal reduction and skipping connection are combined to learn the deep and shallow features of temporal data. Finally, to improve the stability of network training, a mixed loss function is designed to optimize the network training process. Results: We compare our method with different state-of-art algorithms, both visually and with respect to objective quality metrics. The experiments show that our algorithm compares favorably to other state-of-art methods. Conclusions: The proposed BC-DCNN framework provides a new solution to the problem of low-definition ultrasound diagnosis, and successfully solve the ultrasonic image denoising problem while preserving the detailed features.