Research On Weakly Supervision Semantic Segmentation Method For Thyroid Ultrasound Images Based On Semantic Mining

Medical image segmentation technology based on deep learning has greatly promoted the development of computer-aided diagnosis.The difficulty of pixel-level label required by fully-supervised semantic segmentation leads to high data acquisition costs.In contrast,weakly-supervised semantic segmentation requires image-level label and easy data acquisition,which has attracted more and more researchers’ attention.The existing weakly-supervised semantic segmentation technology has developed rapidly,but there are still problems such as sparse initial seeds,incomplete segmentation,and incorrect segmentation.In response to the above-mentioned problems,we carried out research on weakly-supervised semantic segmentation method of thyroid ultrasound images.In order to solve the problem of initial seeds sparseness in weakly-supervised semantic segmentation methods,we propose a weakly-supervised object localization algorithm based on superpixel to provide higher-quality initial seeds for weakly-supervised semantic segmentation.Firstly add atrous space pyramid pooling to the classification network to learn more complete features.Secondly,superpixel optimization module is proposed to increase the prediction value of low-response object pixels by calculating the average value of the superpixel,and alleviate the problem of initial seeds sparseness.In order to solve the problems of incomplete segmentation and incorrect segmentation in weakly-supervised semantic segmentation methods,we propose a weaklysupervised segmentation method for cross-image semantic mining.Firstly we propose the affine self-optimization module,which calculates the correlation between different feature maps before and after the image affine transformation,reduces the difference between different features,and obtains a more complete range of lesions.Secondly,background mining module using the similarity of human tissue structure in ultrasound images is proposed,and the problem of background incorrect segmentation is alleviated by mining similar semantic features between different classes of images.Finally,a multi-feature fusion module is proposed to fuse the shallow and deep features to improve the accuracy of boundary segmentation.We cascades the aforementioned two stages into a weakly-supervised segmentation method for semantic mining combined with superpixel.The segmentation accuracy m Io U and Dice coefficients of this method on the thyroid ultrasound image data set reach 55.41% and 71.29%,respectively,which are higher than the existing weaklysupervised segmentation methods.