Research On Segmentation And Classification Of Ultrasonic Images Based On Self-Supervised Learning

Deep neural networks have improved universal machine learning tasks due to their excellent ability to learn rich semantic features from high-dimensional data.However,large manually labeled datasets are usually required for machine learning.These datasets are often expensive and impractical to obtain due to the time required by professionals to label the data and data protection issues.Selfsupervised learning is an unsupervised learning method,which uses the relevant information of the input data itself as the supervisory signal,has higher data efficiency and stronger generalization ability,and can solve the problem of fewer labeled samples.In this paper,a classification and segmentation algorithm for ultrasonic images is proposed based on self-supervised learning.The research work and achievements are as follows:1.A self-supervised learning method based on mask reconstruction is proposed to achieve ultrasonic image segmentation.Through the pretext task of mask reconstruction,the original ultrasound image is overlaid with multiple evenly distributed mask blocks,and the self-supervised learning is completed by training the model to reconstruct the original image,and the semantic features can be learned with unlabeled data.At the same time,a new segmentation network is proposed to solve the locality problem of convolutional neural networks.We use Transformer encoder and U-Net encoder to extract features from ultrasonic images together,and then use channel attention to fuse them,which can improve the segmentation ability of the model and improve the segmentation accuracy.Experimental results on thyroid and breast ultrasound image datasets show that our algorithm achieved a higher segmentation accuracy with less labeled data for thyroid datasets.In particular,good results are still achieved in the transfer learning experiment of breast images,which proves the effectiveness and generalization of our algorithm.2.A self-supervised learning method based on knowledge distillation is proposed to achieve ultrasonic image segmentation.The method adopts the response-based knowledge distillation method,which takes the segmentation results of the teacher network on ultrasonic images as the target knowledge,calculates the KL divergence as loss to complete the knowledge transfer to a smaller student network.After that,a small number of labeled samples were used for fine tuning to complete the identification of diseased areas.Experimental results on two ultrasonic image datasets show that the proposed method has better performance.3.A self-supervised learning method based on rotation angle prediction is proposed to complete ultrasonic image classification.The pretext task of image classification is constructed by predicting the rotation angle,and the relevant semantic features are learned with unlabeled data.The AMEfficient Net model is also proposed,the attention map extracted from the ultrasonic images by the Vi T is regarded as the feature engineering.The two Efficient Nets are used to extract the original ultrasonic images and its attention maps respectively,and the two features are used comprehensively to classify the benign and malignant diseases,which can improve the classification capability of the model and achieve better classification accuracy.Experimental results on two ultrasonic image datasets show that the AM-Efficient Net proposed in this paper explores the value of unlabeled data,achieving higher classification accuracy with fewer labeled data.