Segmentation of Nasopharyngeal Carcinoma Using 3D CNN Based on Multi-scale Feature Pyramid

Abstract

The invention relates to a nasopharyngeal tumor image segmentation technology in the image segmentation field, in particular to a 3D CNN nasopharyngeal carcinoma segmentation method based on a multi-scale feature pyramid. For the training samples, the experienced radiologist should label several cases of nasopharyngeal carcinoma, use the whole three-dimensional MRI image to establish the data set,and preprocess the data set, then train the training data set with network to obtain high-precision segmentation model. For new cases, the segmentation model can be used to segment MRI images. Compared with the traditional methods, except for manual labeling in the training phase, the other parts can be processed automatically, which greatly reduces the demand for experienced physicians and achieves higher accuracy compared with the five mainstream networks.

Description

technical field [0001] The invention relates to a nasopharyngeal tumor image segmentation technology in the field of image segmentation, in particular to a 3D CNN nasopharyngeal cancer segmentation method based on a multi-scale feature pyramid. Background technique [0002] Nasopharyngeal carcinoma is one of the most common cancers in the nasopharyngeal tissue structure, and it often occurs in malignant tumors on the roof and side walls of the nasopharyngeal cavity. The disease is one of the malignant tumors with high incidence in the world, and its incidence rate is the first among malignant tumors of the ear, nose and throat, accounting for 1% of all cancer diseases. Nasopharyngeal tumors have a more complex anatomy compared to other types of tumors such as lung tumors. It is spatially similar to several tissues (air, bone, muscle, and mucous membrane) that process similar image intensities, and the shape and size of nasopharyngeal carcinoma, as well as non-uniform tumor ...

AI Technical Summary

Problems solved by technology

This processing method will cause two problems: a large number of redundant calculations lead to low time efficiency, and DCNN cannot learn global features

In summary, there are two problems with the original method. On the one hand, there is a lack of reuse of original features. On the other hand, only fusion in the upsampling phase does not promote rapid convergence in the upsampling phase.

Images