Article
Observer-independent automated analysis of label-free multiphoton images of human brain tumors
Search Medline for
Authors
Published: | June 9, 2017 |
---|
Outline
Text
Objective: Intraoperative fast and reliable recognition of tumor borders and diagnosis is a challenge in neurosurgery. Label-free multiphoton imaging of brain tumors retrieves morphochemical information with subcellular resolution that allows detailed tissue analysis and was suggested for intraoperative application. Therefore, we tested different settings to identify parameters that balance minimal acquisition time and sufficient image quality for automated texture analysis to discriminate relevant tissue types.
Methods: Samples of human brain tissue was obtained during routine surgery and cryosections were prepared. We analyzed glioblastoma, gray matter, white matter and necrotic tissue (n=10). Based on the comparison with reference HE stainings, five representative areas were selected on each sample and subjected to multiphoton imaging: CARS (coherent anti-Stokes Raman scattering) TPEF (endogenous two photon excited fluorescence) and SHG (second harmonic generation) were simultaneously acquired. For each position, pixel size was set to 0.2, 0.5 and 1 µm and averaging was changed between 1, 4 and 16 resulting in acquisition times from 200 ms to 16 s. In total, 1800 images were obtained. Gray values of each modality were normalized (min to max) and texture analysis was performed using Matlab functions.
Results: : Images of white matter were dominated by axons (CARS), gray matter displayed cell bodies with fluorescent inclusions (TPEF) while glioblastoma had a homogenous appearance. Necrotic areas exhibited large lipid droplets (CARS). First order parameter (mean, standard deviation, kurtosis, skewness, entropy) and second order parameters (contrast, correlation, energy, homogeneity, for 8 and 22 µm distances) were calculated separately for each channel. The comparison of texture parameters of images obtained with different acquisition settings showed that an averaging of 4 µm and a pixel dimension of 1 µm is sufficient for reliable discrimination. For instance, glioblastoma and white matter showed significant differences for CARS: standard deviation, skewness and entropy, for TPEF: standard deviation, kurtosis, skewness and entropy and for SHG: mean and skewness (t-test, P<0.05). The second order texture parameters of CARS and TPEF likewise provided to be reliable discriminators of tissue types.
Conclusion: CARS-TPEF-SHG images of intermediate signal-to-noise ratio and a lateral resolution of 1 µm are sufficient to extract tissue characteristics using automated texture analysis. Therefore, observer-independent, reliable tissue classification of glioblastoma, white and gray matter as well as necrotic areas can be obtained. Based on this objective analysis, intraoperative multiphoton images could be classified in real-time to provide this information immediately to the surgeon. The current approach is extended to additional pathologies and tissue types.