Thyroid nodule semi-supervised segmentation method based on attention mechanism

Abstract

The invention discloses a thyroid nodule semi-supervised segmentation method based on an attention mechanism. The method comprises the following steps of: 1, carrying out preprocessing of a thyroid ultrasonic image, and removing an edge information region in the image; 2, constructing a semi-supervised segmentation neural network, performing classification and segmentation prediction tasks on theultrasonic image, and adjusting a network structure to adapt to a specific application scene; 3, adding an attention mechanism into the semi-supervised segmentation neural network to improve the network effect; 4, measuring the performances of a semi-supervised segmentation algorithm and an existing full-supervised segmentation algorithm in the field of thyroid nodule auxiliary diagnosis through an intersection-parallel ratio and a Dice coefficient; and 5, continuously reducing the number of the pixel-level labels, and observing the change condition of the network performance. According to theinvention, the thyroid nodule semi-supervised segmentation method based on an attention mechanism benefits from the semi-supervised effect of a small number of pixel-level labels while keeping the high segmentation performance of the semi-supervised segmentation model, learns the real benign and malignant characteristics of the nodules and improves the benign and malignant classification capacity.

Description

technical field [0001] The invention belongs to the fields of deep learning, computer-aided medical treatment and medical image processing, and relates to neural network classification technology and semi-supervised learning technology, especially a semi-supervised segmentation method for thyroid nodules based on an attention mechanism. Background technique [0002] The research and application of medical aided diagnosis using deep learning methods covers skin diseases, brain diseases, lung inflammation, and thyroid nodules. Jinlian Ma et al. [1] first used convolutional neural networks in the ultrasound diagnosis of thyroid nodules in 2017. They fine-tuned two networks pre-trained in the ImageNet database by training separately, and achieved a diagnostic accuracy of 83.02% ± 0.72% by concatenating feature maps. However, its pre-trained network is based on ImageNet, a natural scene image dataset, and its pre-trained features are mostly natural scenes, not pathological featu...

AI Technical Summary

Problems solved by technology

However, the above methods rely either on the combination of attention maps trained by one network for different erasure steps, or the combination of attention maps of different networks, and the attention of a single network is still only focused on the most discriminative regions.

In addition, because the proportion of the lesion area in the medical image is too small compared with the normal tissue area, the influence of the background on the model cannot be ignored, that is, the weak supervision method cannot ensure the learning accuracy of the model

Images