Oncolohy Medical Image Semanyic Segmentation And Classification Based On Multitask Learning

Breast cancer is one of the most common and serious cancers among women worldwide.The incidence has been on the rise since the 1970 s and has become the most diagnosed female cancer in the world.In traditional ultrasound examinations,doctors use their own professional knowledge and clinical experience to diagnose according to ultrasound images.Subjectively,they will be interfered by other body tissues and shadows,so there will be missed diagnosis or misdiagnosis,so clinical diagnosis The need to introduce computer-aided design can reduce human intervention and improve diagnostic performance.Most computer-aided design systems today use deep learning for medical image processing.Most deep learning algorithms require a large number of data sets to support the training of the model in order to get better results.Due to the particularity of the ultrasound image of the breast,it can be obtained.The sample is difficult to meet the hard requirements of the deep learning algorithm,so the general deep learning method can not achieve good results when performing breast ultrasound image processing.Therefore,this paper mainly studies how to use the multi-task learning algorithm to achieve better semantic segmentation and classification of breast ultrasound images in the case of fewer samples,so as to better assist physicians in the diagnosis of breast cancer.This paper first analyzes the research status of medical ultrasound image segmentation and classification at home and abroad,compares the advantages and disadvantages of each algorithm,and proposes the idea of ??breast ultrasound image classification and semantic segmentation based on multi-task learning.Convolutional neural network is good at image level.The classification of this method is used to judge the benign and malignant breast tumors.U-net network can realize pixel-level classification,so it is used to semantically segment breast tumor images,and combine these two algorithms and multi-task learning ideas.Two multi-task models are used to improve the effects of benign and malignant discrimination and semantic segmentation through weight sharing.The two models use different methods of sharing weights for training.For the multi-task model,a loss function that can adjust the proportion of classification and semantic segmentation is proposed.By adjusting the proportion of two tasks,one of the tasks is too large,and the impact is too large.The learning effect of another task achieves better experimental results.At the end of this paper,a separate confounding neural network based breast ultrasound image benign and malignant discrimination model,U-net based semantic segmentation model,and two multi-task models proposed in this paper are experimentally analyzed.The experimental results show that the proposed The multi-task learning model can improve the classification of convolutional neural networks and the semanticsegmentation effect of U-net networks.In the case of small samples,the sample information can be fully exploited to improve the effect of model classification and semantic segmentation.