In‐vivo multispectral video endoscopy towards in‐vivo hyperspectral video endoscopy

Abstract

For in‐vivo diagnostics of cancer and pre‐cancer in the stomach, there is no endoscopic procedure offering both high sensitivity and high specificity. Our data suggest that multispectral or hyperspectral imaging may be helpful to solve this problem. It is successfully applied to the detection and analysis of easily reachable carcinomas, ex‐vivo samples of hollow organ mucosal carcinomas and also histological samples. An endoscopy system which allows flexible multispectral videoendoscopy for in‐vivo diagnostics has so far been unavailable. To overcome this problem, we modified a standard Olympus endoscopy system to conduct in‐vivo multispectral imaging of the upper GI tract. The pilot study is performed on 14 patients with adeno carcinomas in the stomach. For analysis, Support Vector Machine with linear and Gaussian Kernel, AdaBoost, RobustBoost and Random‐Forest‐walk are used and compared for the data classification with a leave‐one‐out strategy. The margin of the carcinoma for the training of the classifier is drawn by expert‐labeling. The cancer findings are cross‐checked by biopsies. We expect that the present study will help to improve the further development of hyperspectral endoscopy and to overcome some of the problems to be faced in this process.

Factors influencing the accuracy for tissue classification in multi spectral in-vivo endoscopy for the upper gastro-internal tract

“…However due the usage of the Gaussian filter, the noise is expected to be higher. Stronger noise reduction or more photons are not needed due to the fact that the noise level is similar or even better than in our previous in-vivo study 21 . Using the Gaussian filtering instead of decreasing the spatial resolution of the detector has two advantages: First, the resolution is similar to at least older endoscopes.…”

“…However, the transfer of these results to in-vivo by our previous publications showed significant worse results. An accuracy of 64% is reached in the first study 21 and an accuracy of 68% in the second study 22 . These improvements could be reached by using elaborated methods for data analysis.…”

“…To derive the optical properties and include inter patient variations, the following strategy is chosen: First, a centre value is calculated and, second, the variations are derived. The centre value of the optical properties are derived by averaging the optical properties for each spectral bandwidth, using the spectra of the illumination source as weight from our older clinical study for carcinomas in the stomach 21 . This is done for each wavelength band to ensure that values match our the previous publication 21 .…”

“…The centre value of the optical properties are derived by averaging the optical properties for each spectral bandwidth, using the spectra of the illumination source as weight from our older clinical study for carcinomas in the stomach 21 . This is done for each wavelength band to ensure that values match our the previous publication 21 . By integrating the optical properties abbreviated with OP times the normalized intensity (I) of the light source, the final optical properties are calculated as weighted mean:…”

“…One way of spectroscopic measurement is multi/hyper spectral imaging (MSI/HSI). So far, both have been proven successful in finding cancer in different parts of the human body such as: larynx 5 with a sensitivity of 55-87% and specificity of nearly 100% 6 , cell analysis from cervix 7 with a sensitivity ranging from 66 to 100%, breast 8,9 , colon [10][11][12][13][14][15][16][17] or oesophagus/stomach [18][19][20][21][22] . Thereby, the group from Sakaida reached an accuracy of 79% in the first study 19 and 85% in their second study 20 in an ex-vivo setting.…”