Nasopharyngeal carcinoma focus image segmentation device, equipment and computer readable storage medium

Abstract

The invention discloses a nasopharyngeal carcinoma focus image segmentation device, and relates to the field of medical image processing, and the device comprises an image obtaining module which is used for obtaining a to-be-segmented image; a feature extraction module which is used for extracting image features of the to-be-segmented image; a neural network module which comprises a neural networkused for carrying out image segmentation on the to-be-segmented image according to the image features; the neural network further comprises a position acquisition unit and a semantic segmentation unit; the position acquisition unit is used for acquiring position information of a focus in the to-be-segmented image according to the image features; and the semantic segmentation unit is used for exciting a feature map contributing to segmentation according to the image features in combination with the position information so as to obtain a semantic segmentation result. The embodiment of the invention further provides nasopharyngeal carcinoma focus image segmentation equipment and a computer readable storage medium, a neural network model can be effectively simplified, and the segmentation precision of nasopharyngeal carcinoma focus image segmentation is improved.

Description

technical field [0001] The present invention relates to the field of medical image processing, in particular to a lesion image segmentation device, equipment and computer-readable storage medium for nasopharyngeal carcinoma. Background technique [0002] Nasopharyngeal carcinoma is one of the high incidence malignant tumors, and its incidence rate is the first among ear, nose and throat malignant tumors. In order to treat patients with nasopharyngeal carcinoma, it is necessary to identify nasopharyngeal carcinoma lesions. In clinical practice, image processing of nuclear magnetic resonance images is generally used to identify nasopharyngeal carcinoma lesions. [0003] In the existing medical image processing, commonly used image segmentation algorithms include threshold-based segmentation methods, edge-based segmentation methods, region-based segmentation methods, cluster analysis-based image segmentation methods, wavelet transform-based segmentation methods, Segmentation m...

AI Technical Summary

Problems solved by technology

[0004] However, the inventor found in the research that the image segmentation method based on edge and threshold is too simple, does not make good use of the spatial information of pixels, the segmentation result is extremely susceptible to noise interference, and often has broken edges, which require post-processing. This method requires segmentation The object color texture is relatively compact, and the intra-class variance is small, which is only suitable for text image processing, such as license plate and fingerprint; the segmentation method based on region (growth, split) often causes over-segmentation, that is, the image is divided into too many regions, and as An iterative method, the space and time overhead are relatively large, noise and grayscale inhomogeneity may cause holes and over-segmentation, and it is often not very good at processing shadow effects in images

The image segmentation method based on cluster analysis does not consider spatial information, and is sensitive to noise and gray level inhomogeneity. Clustering needs to determine the number of classes; the segmentation method of wavelet transform needs to select a suitable filter; it is difficult to avoid it after processing based on mathematical morphology methods. There will still be a large number of short lines and isolated points that do not match the target, so post-processing is required in addition to pre-processing; what network structure needs to be selected based on the artificial neural network method requires a large amount of data, slow speed, and complex structure; The fitness function also needs to determine the determination of the crossover probability, and it may also converge to a local optimum

The method based on deep learning is mainly based on the segmentation method of convolutional neural network. The existing semantic segmentation model is mainly aimed at the segmentation of general images, while the medical image of nasopharyngeal carcinoma has more complex image content, which is aimed at the segmentation target is small and has invasiveness. Wetness, the contrast and texture of normal tissue are very similar, and the anatomical structure is complex. Even a segmentation model suitable for medical image segmentation such as Unet will appear in the task of primary tumor segmentation of nasopharyngeal carcinoma. Some problems, such as 1) false positives are high, 2) the edge of the segmentation result is not clear; in addition, for nasopharyngeal carcinoma data, there are typical segmentation methods; 3) multimodal image data cannot be directly processed, and nasopharyngeal carcinoma images are generally There are multiple modal data, such as T1-weight, CET1-weight, T2-weight imaging images; 4) There are still relatively few methods for the segmentation of primary tumors of nasopharyngeal carcinoma based on deep convolutional neural networks

Images