Cancer hyperspectral image segmentation method and system based on double-branch attention deep learning

Abstract

The invention discloses a cancer hyperspectral image segmentation method based on double-branch attention deep learning. The method comprises the steps: firstly carrying out the preprocessing of a cancer hyperspectral image, and dividing a data set into a training set and a test set; then constructing a double-branch deep convolutional neural network based on an attention mechanism, and training the network through the divided training set; and then the trained neural network is used to test the test set, and finally the purpose of segmenting the cancer hyperspectral image is achieved. According to the method, a double-branch structure is constructed according to the characteristics of the cancer hyperspectral image, spatial information and spectral information of the hyperspectral image are extracted respectively and fused, and finally a prediction result is obtained. The invention further discloses a cancer hyperspectral image segmentation system based on double-branch attention deeplearning.

Description

technical field [0001] The present invention relates to the technical fields of hyperspectral image processing and medical image processing, in particular to a cancer hyperspectral image segmentation method and system based on dual-branch attention deep learning. Background technique [0002] Carcinoma refers to malignant tumors originating from epithelial tissue, and is the most common type of malignant tumors. Most types of cancer lack typical clinical manifestations and diagnostic methods in the early stage. When patients have typical symptoms, they have basically entered the middle and late stages and lost the opportunity for radical surgery. There are many diagnostic methods for cancer, mainly including clinical diagnosis, physical and chemical diagnosis, surgical diagnosis, cytopathological diagnosis, and histopathological diagnosis. Histopathological diagnosis can understand the benign and malignant nature of the tumor, judge the prognosis of cancer, provide direct h...

AI Technical Summary

Problems solved by technology

[0006] Although the existing hyperspectral image processing methods can achieve hyperspectral image segmentation, there are still three deficiencies in the processing flow: First, the existing methods cannot extract hyperspectral images on multiple scales when extracting hyperspectral features. Spatial attention information and channel attention information of

The commonly used attention mechanism often only extracts features on a single scale, but there is no corresponding strategy for hyperspectral images with different scale features

Second, the existing network structure cannot extract image features on a larger spatial scale.

However, cancer hyperspectral images have large-scale morphological features due to factors such as tissue or cell structure and shape, so the method of using image blocks as network input is not applicable.

Third, in the case of using traditional machine learning methods, the processing flow is often not an end-to-end model, resulting in loss of accuracy of results and complicated processing steps

SVM is a commonly used machine learning method. However, when dealing with hyperspectral images, it is often necessary to first perform dimensionality reduction operations on the data through the PCA method. However, since the model is not an end-to-end structure, the dimensionality reduction strategy of PCA and the method of segmenting hyperspectral images The goals are different, so some important information that is helpful for segmentation may be lost during the dimensionality reduction process, and the entire model cannot be optimized

These three deficiencies will lead to less robustness and lower efficiency in practical hyperspectral image segmentation problems

Images