A Novel Method to Denoise CEUS Image Combining Bidirectional ConvLSTM with 3D DnCNN

Due to the limitation of imaging technology, there is a lot of speckle noise in contrast-enhanced ultrasound (CEUS) images, which affects the improvement of resolution and the correct diagnosis made by doctors. It is difficult for existing methods to balance better image denoising and more detailed preservation. This paper presents a network combining bidirectional ConvLSTM (Convolutional Long Short-Term Memory) with 3D DnCNN to generate clearer images which we name BCDCNN. This makes full use of the correlation between CEUS video frame sequences. Firstly, we design an abnormal frame detection module to filter out abnormal frames from the video and combine several CEUS videos with synthetic CEUS videos to avoid over-fitting and improve the generalization ability of the model. Secondly, a three-dimensional convolution operator is imported to improve the two-dimensional DnCNN denoising model, so as to process image sequences on the time dimension. Finally, the bidirectional ConvLSTM is imported to learn the time characteristics of video, which makes full use of the inter-frame information. By observing the visual quality and various quantification indicators, our method performs better in denoising and detail preservation than the current mainstream methods, which is conducive to future medical analysis and diagnosis.