Ultrasonic image-based thyroid nodule automatic segmentation method

Abstract

The invention relates to an ultrasonic image-based thyroid nodule automatic segmentation method. The method comprises the following steps of 1, establishing a training set and a test set by using a public thyroid nodule data set; processing an image and an XML file, performing cutting to obtain the image and a label corresponding to the image, and dividing the image into a training set, a test set and a cross validation set in proportion; 2, establishing a thyroid nodule segmentation network, wherein network structures such as cavity convolution, dense connection and repair-module (correction module) are used; 3, training the thyroid nodule segmentation network by using the training set to obtain a trained segmentation model; and 4, verifying the segmentation model in a test set, and segmenting the thyroid nodule to obtain nodule lesion information. The thyroid nodule segmentation network is trained by using the training set, the trained segmentation model is obtained, and doctors are assisted in improving the diagnosis efficiency.

Description

technical field [0001] The invention relates to an automatic segmentation method for thyroid nodules based on ultrasonic images, and relates to the field of computer-aided diagnosis of medical images. [0002] technical background [0003] The thyroid is the largest endocrine organ in the human body and an extremely important organ that regulates metabolism in the human body. Due to the influence of factors such as people's eating habits, irregular living habits, and high mental stress, the incidence of thyroid disease is increasing year by year. The incidence rate of thyroid disease is more than 50%, and all age groups have the possibility of onset. The incidence of thyroid nodules is high, but more than 90% of them are benign nodules, and drug treatment is very effective for malignant nodules. Therefore, as long as it is discovered by medical examination in the early stage and clinical treatment is carried out as early as possible, the cure rate can be increased to more t...

AI Technical Summary

Problems solved by technology

[0007] 1) Due to the fuzzy ultrasound images of thyroid nodules, it is a difficult segmentation task in medical image processing. The performance of existing segmentation algorithms in thyroid nodule segmentation is still not very satisfactory. Some researchers use hollow convolution to increase the Receptive field, the effect has been improved, but the segmentation algorithm using hole convolution has a high false negative rate, and there will be more obvious holes

[0008] 2) Most of the existing segmentation algorithms are derived from natural image segmentation, and have not been optimized for the characteristics of thyroid images, especially the necessary lightweight processing. The network structure is relatively complex, so the speed of training and testing is relatively slow

[0009] 3) The existing open source thyroid nodule data sets are generally small, and the batch normalization (Batch Normalization) adopted in the existing segmentation algorithm is not as ideal when the batch size is small as when the batch size is large, resulting in segmentation Algorithms don't perform as well as they should

[0010] 4) The loss function of existing segmentation algorithms is usually cross entropy, which is a general loss function, but it is not the most suitable for thyroid nodule segmentation

Images