Deep Learning-based Segmentation And Reconstruction Of Plane-wave B-mode Ultrasound Images

In medical imaging applications,plane-wave B-mode ultrasound imaging has become the primary choice for physicians and patients in the imaging process due to its non-invasive and inexpensive features.However,while physicians require high accuracy in the segmentation of medical images,ultrasound imaging technology is limited by echo interference from in vivo tissues.It has inherent problems such as low resolution and blurred edge details.This undoubtedly makes it more difficult for physicians to segment lesions manually.Therefore,there is an urgent need to find a solution to improve image resolution and assist physicians in segmenting lesions.Based on deep learning theory,this project designs a series of reconstruction and segmentation schemes for plane-wave B-mode ultrasound images.For the characteristics of blurred edges and poor clarity of plane-wave B-mode ultrasound images,super-resolution reconstruction is used to improve image clarity and reduce the difficulty of physicians’ work.And for medical targets of different scales,different segmentation models are proposed,which in turn assist physicians in completing the segmentation task.The main research work of this thesis is as follows:(1)A segmentation model incorporating multi-scale features is proposed for small target region images represented by brachial plexus nerve ultrasound images,which are improved based on Feature Pyramid Network(FPN)structure.The design uses a multi-scale feature extraction module and prediction using the improved bidirectional FPN structure.The model achieves a Dice performance of 70.28% on the Kaggle public dataset,significantly improving over the mainstream methods.(2)A segmentation model based on Pyramid Attention Network(PAN)structure(AMS-PAN)is proposed for the characteristics of large target unstable regions represented by breast ultrasound images.The attention module is improved and added based on PAN,and multi-scale perceptual field extraction features are designed.The Dice of this model on BUSI and OASBUD datasets are 80.71% and 79.62%,respectively,with a significant improvement compared with current advanced methods.(3)A GAN model for single-image super-resolution reconstruction task(SISR)is proposed for the problem of low contrast and poor definition of plane-wave ultrasound images.A full-scale jump-connected U-Net with an attention mechanism is presented as a generator,and an improved Patch discriminator is used to complete the adversarial process.A combined loss function is designed for the model and trained on the PICMUS dataset to enhance the quality of plane-wave ultrasound images.The reconstruction of the BUSI dataset verifies its effectiveness in improving segmentation accuracy.(4)A sampling loss-free GAN(WSRGAN)is proposed for the problem that sampling loss in the reconstruction process is challenging to avoid.The sampling model with Discrete Wavelet Transform(DWT)and Inverse Discrete Wavelet Transform(IDWT)instead of pooling and sampling layers in deep learning is designed.The adaptive wavelet attention mechanism and wavelet counteracting loss function are proposed based on the results of the wavelet decomposition.Experiments on the PICMUS dataset verify the effectiveness of each module,and its practical application in segmentation tasks is demonstrated in experiments on the BUSI dataset.