Ultrasonic Image Assisted Diagnosis Of Breast Tumors Based On Feature Fusion And Attention Mechanism

In 2020,breast cancer will be the most diagnosed cancer in women worldwide,posing a serious health risk to women.Computer-aided diagnostic systems based on breast ultrasound images can give an important "second opinion".Earlier computer-aided diagnosis systems required doctors to mark manually,which was time-consuming and laborious.With the rapid development of deep learning,tumor segmentation and tumor classification techniques have achieved excellent performance.The segmentation of lesion regions by tumor segmentation technology and the classification and identification of lesion images by tumor classification technology have significantly improved the efficiency of clinical breast cancer diagnosis.However,due to the image quality of ultrasound images,computer-aided diagnosis techniques still have the problems of low recognition rate and slow recognition.This paper is dedicated to solving the problems and difficulties of segmentation and classification of breast tumor lesion regions by using deep learning methods,and has conducted an indepth study.The main research contents are as follows.(1)Research on automatic tumor segmentation algorithms for breast ultrasound image research.For the current mainstream segmentation models to extract features with low efficiency and easy to cause feature information loss,this paper proposes an encoding-decoding lesion segmentation network based on Mobile Net backbone and feature fusion,firstly,the encoder extracts the multi-scale features of the image to expand the network perceptual field,then the decoder fuses the feature map and then up samples the image semantic information to restore the image semantic information,and designs the loss function Constrain the image focal region and background region to assist the model for training.The model is able to achieve Mean Io U of 0.791 and Dice coefficient value of 0.855 on the public dataset.(2)In order to better link tumor segmentation and tumor classification together to form a complete auxiliary diagnostic scheme,segmentation enhancement preprocessing is performed on breast ultrasound images using the a priori information of tumor segmentation to achieve the effect of noise removal,detail and edge enhancement on the original to generate RGB images for tumor classification.(3)To study automatic tumor classification algorithms for breast ultrasound image research.Improving the design of lightweight visual attention network VAN for tumor classification,introducing hybrid feedforward neural network module,extracting image local and detail information using large kernel attention mechanism,and classifying image features using Softmax classifier.Experimental validation was performed on a private dataset for tumor classification,and Accuracy and F1-score reached 0.9066 and 0.9079,respectively.Finally,the contribution of image preprocessing strategy was analyzed to prove the effectiveness and superiority of the proposed breast ultrasound image assisted diagnosis scheme.