Study Of Medical Hyperspectral Image Lesion Localization Based On Deep Learning And Space-Spectrum Combination

In recent years,hyperspectral imaging technology has developed rapidly and achieved good results in agriculture,military,geology,medical and other fields.At present,medical hyperspectral imaging technology,which combines clinical diagnosis with hyperspectral technology,has become a new research trend.Based on medical hyperspectral imaging technology,this thesis aims to better extract the deep space spectrum features inside medical hyperspectral data by building a deep learning model,in order to provide more auxiliary information for pathological diagnosis of surgeons.Taking glioblastoma tissue slices as an example,a combined classification model based on mixed network and attention mechanism was proposed respectively,which will produce important research value for the diagnosis of brain cancer in the future.The research of this thesis consists of four main areas:(1)For the noise generated by the system during the acquisition of medical hyperspectral images,the high dimensionality of the data and the difference in the radiation intensity of each pixel caused by the inhomogeneity of the brain surface,pre-processing methods such as image calibration,dimensionality reduction based on principal component analysis and data normalization were used to improve the signal-to-noise ratio of the data and weaken the interference of noise,respectively.(2)In order to fully extract the spectral and spatial information of medical hyperspectral images and ensure the effective use of spatial features at multiple scales,this thesis proposes a hybrid network-based hyperspectral image classification of glioblastoma(Hy SSNET)model,which overcomes the defects of using two-dimensional convolutional neural networks and three-dimensional convolutional neural networks alone,reduces the computational complexity,and achieves the accurate classification of glioblastoma.(3)In the actual research process,the classification accuracy is often low due to the scarcity of sample data.To solve this problem and achieve accurate identification of glioblastoma under small samples,this thesis improves the hybrid network-based classification method and proposes the attention mechanism-based glioblastoma classification and diagnosis(SCBAM-SA-Hy SSNET)model,which extracts more valuable spatial spectral information inside medical hyperspectral images and obtains higher classification accuracy by adding the random convolutional attention mechanism(SCBAM)and spatial attention mechanism modules.(4)A hyperspectral image classification system for glioblastoma combined with deep learning techniques was designed.By encapsulating the network with good classification performance trained in advance and then selecting the glioblastoma data for processing,the classification result maps under different network models can be obtained,and the whole process is labor-saving and efficient,which facilitates the identification of tumors for surgeons.