Research And Application Of Ultrasound Medical Image Intelligent Aided Diagnosis Algorithm

Objective: Ultrasound equipment has the advantages of easy operation,safety,rapidity,and real-time imaging,and is widely used in clinical diagnosis in various tissues and organs.Automatic diagnosis of ultrasound images can effectively improve the efficiency of physicians’ interpretation and reduce the impact of subjective interpretation.Currently,deep learning-based intelligent image processing algorithms are excelling in computer vision tasks such as medical image processing,achieving a qualitative leap in the accuracy of some tasks and providing more possibilities for truly intelligent diagnosis.Ultrasound medicine is widely used in clinical applications,and this study is based on two tasks: ultrasound image assisted diagnosis of breast cancer and fetal head circumference measurement.First,at the algorithmic level,it is difficult to achieve efficient and robust detection and segmentation algorithms by traditional image processing methods due to the unique nature of ultrasound images.Deep learning plays an increasingly important role in the research of computer-aided diagnosis algorithms by deeply extracting the features of the lesion region of ultrasound images through the unique hierarchical structure,which improves the accuracy of detection and segmentation.Therefore,we delve into the problem of lesion region segmentation and benign-malignant classification of breast cancer ultrasound images and propose a multi-task deep learning algorithm based on different attention mechanisms,and for the problem of fetal head circumference measurement,we propose a lightweight fetal head circumference auto-regression measurement algorithm.Secondly,considering the practical application of the algorithmic model,the existing method of deploying remote access based cloud-based deep learning models improves the accuracy and availability of detection results,but the time delay of inference results is not negligible,which limits the application of deep learning in ultrasound.Therefore,we optimize the proposed lightweight fetal head circumference measurement model comprehensively in terms of time and accuracy,and implement the model for mobile application deployment.Methods and results: In this study,we propose a multi-task model for segmentation and classification of breast cancer ultrasound images based on the hypothesis that more attention to lesion regions can effectively improve the discriminative ability for lesion types.A multi-task model based on different attention mechanisms is proposed for breast cancer ultrasound image segmentation and classification;a deep convolutional neural network is used for automatic feature extraction and end-to-end regression of fetal head ellipse localization and head circumference measurement,and the trained lightweight model is deployed to mobile device to implement real-time processing of fetal head ellipse localization and head circumference measurement.(1)The imaging process of ultrasound medical images is prone to noise and the images are generally complex,and currently the clinic mainly relies on manual interpretation.Intelligent processing of ultrasound medical images plays an important role in assisting clinicians in diagnosis,reducing clinicians’ workload and improving the objectivity of diagnosis.In this study,we propose a multi-task learning(SHA-MTL)model for simultaneous segmentation and classification of breast cancer ultrasound images based on soft-hard attention,whose backbone network is composed of dense CNN encoder and upsampling decoder,the encoder network is mainly responsible for the extraction of classification features,the upsampling decoder is the segmentation branch,and the decoder and encoder are connected by attention-gated(AG)units with soft attention mechanism to implement the fusion training of local and global features in the network.The model performs cross-validation experiments on a public ultrasound breast dataset with category and mask labels and performs a comprehensive analysis of the two tasks.For the segmentation task,the SHA-MTL model improves the Sen.and Dice values by 2.27% and 1.19%,respectively,and improves the classification accuracy and F1 values by 2.45% and 3.82%,respectively,compared to the single-task model.The experimental results verified the validity of our model and showed that paying more attention to lesion regions through the attention mechanism can effectively improve the model’s ability to discriminate lesion types.(2)At present,portable ultrasound devices are spreading rapidly and covering more application scenarios,and new requirements for intelligent aided diagnostic algorithms have been put forward.As an important physiological indicator during pregnancy,clinical measurement of fetal head circumference mainly relies on manual measurement by physicians.Deep learning based intelligent algorithms have good feature extraction performance and can be designed for task-specific end-to-end models to simplify processing steps.This study involves two aspects of work:lightweight algorithm research and deployment of the model on mobile applications.First,we propose an automatic fetal head ellipse detection and circumference measurement model,which is capable of simultaneously detecting the fetal head region and measuring the head circumference on a lightweight network.A large number of validation experiments show that the model obtains the ellipse detection Dice accuracy close to that of the segmentation model and the head circumference estimation can be obtained quickly,while the number of parameters is less than that of the segmentation-based model.Secondly,the proposed model was deployed and evaluated on mobile using Swift language and Core ML framework in this study.The results show that the model can be able to provide highly accurate and real-time fetal head ellipse detection and estimate fetal head circumference in mobile devices.Conclusion: In this paper,by studying ultrasound medical image intelligent assistance algorithms and their applications,we proposed an attention mechanism-based ultrasound medical image intelligent processing model and a fetal head circumference lightweight auto-regression and measurement model,and deployed the latter for mobile applications.The attention-based ultrasound medical image intelligent processing model provides a new idea for medical image lesion region segmentation and lesion type classification by exploring the auxiliary role of attention mechanism on ultrasound medical image multitasking while effectively fusing the local and global features of the two tasks through the attention mechanism.The fetal head circumference lightweight auto-regression and measurement model uses elliptic parametric regression with fewer parameters than the segmentation model to establish the connection between image features and measured biological indicators,and detects fetal head ellipse and estimates head circumference;furthermore,we deploy the model for mobile applications,so that it can achieve real-time automatic fetal head detection and measurement,which can play a role in the research of intelligent ultrasound portable devices.