In vivo tumor wavelength band selection using Hierarchical clustering and PCA with NIR-Hyperspectral Data

Abstract

This paper presents a new method of wavelength selection combined with principle component analysis (PCA) and a hierarchal method for hyperspectral data analysis. Hyperspectral data analysis is a combination of imaging and spectroscopic technology, and is utilized in several fields. In the medical field, if it is possible to distinguish the region of interest by selecting a feature wavelength without the intervention of manufacturers, spectral application as a surgery support system would be feasible. There are several analysis techniques to extract the features using hyperspectral data; however, many of these methods are premised on using all wavelengths. Considering application to endoscopes and other medical devices and reducing number of wavelengths, these methods are not applicable. PCA is a popular analysis for reducing the number of wavelengths, and has been adapted for extracting general and clear features of specimens. However, such extracted features from the body tend to highlight the features of protein and blood making it difficult to extract the features of cancer or other diseases. This paper proposes a two-type feature extraction method that depends on "difference" and "similarity". In this method, the PCA and hierarchal classification analysis are combined to merge the clusters based on similarity. The feasibility of the proposed method is verified by applying the method to discriminating the cancer of mouse: colonized cancer tissue after three weeks and colonized cancer tissue after one week. The results show that the proposed method can extract the two-type feature. The proposed method is had an accuracy of 75.9% in reducing the number of wavelength bands from 256 to 8 for mature cancer in mice. Although less accurate than using all wavelength bands, the eight wavelength bands are distinguished using a simplified convolutional neural network model with an accuracy of 75.9%. The proposed method is able to select the feature wavelength bands for the target. As the learning model for improving the accuracy advances beyond the simple model and the new feature is extracted in the selected wavelength bands, the method is applied to a diagnosis supporting system for endoscopic surgery and more.

Diagnosis of Brain Lesions, Glioma, Multiple-Sclerosis and Metastases from MRI: An efficient classifier-aided method using Refractive Index as a surrogate Biological Marker

“…Clustering of diseases and tissues can also be done using SVMs. It is worth mentioning that machine learning tools like clustering, SVM and neural networks (Principal Component Analysis) have previously been used for similar purposes on MRI data [5,10,[13][14][15]. .…”