The Application Of Deep Learning In Image Segmentation

Thyroid disease is one of the nodular diseases with the highest incidence in the adult population,and its incidence is higher every year.Therefore,how to effectively diagnose thyroid nodules has become more and more important.The emergence of ultrasound examination makes it possible to diagnose thyroid diseases quickly and effectively.ultrasound examination has the advantages of real-time,small damage and repeatability,which has become more and more important in clinical medical diagnosis.However,the contrast of ultrasound medical images collected by medical instruments is relatively low,and thyroid nodules have no fixed shape,which brings certain difficulties to the diagnosis of thyroid disease.Therefore,there is also an urgent need for computer-aided diagnosis technology.This article uses deep learning to segment thyroid ultrasound images.The main research content is how to effectively use neural network models to accurately segment thyroid ultrasound images.There are two main problems in using deep learning to segment thyroid ultrasound images:(1)The data volume of medical images is relatively small,making it difficult to effectively train the network model;(2)Because the thyroid ultrasound image has no fixed shape and the target boundary is relatively blurred,the existing network model is not ideal for its segmentation effect.Therefor,this paper makes adjustments and improvements to design a segmentation method suitable for thyroid ultrasound images.The research contributions and innovations of this article mainly focus on the automatic segmentation of thyroid nodule ultrasound images in clinical medicine.This paper proposes a segmentation method suitable for ultrasound images of thyroid nodules.This method is based on the U-Net network model and combines the feature pyramid module,multi-scale input mechanism and attention mechanism.This method can effectively segment thyroid ultrasound images;In addition,this paper also proposes an improved loss function,which is derived from binary classification,which can effectively balance the accuracy and recall of image segmentation,and has a certain effect on the segmentation of small sample targets.Finally,this article also uses clinical medical thyroid ultrasound images to experimentally verify the proposed segmentation method and loss function,and through comparative analysis with related segmentation models and loss functions,it is proved that this method can effectively segment thyroid ultrasound images.