Dataset Information

Deep Neural Network for Differentiation of Brain Tumor Tissue Displayed by Confocal Laser Endomicroscopy.

ABSTRACT:

Background

Reliable on site classification of resected tumor specimens remains a challenge. Implementation of high-resolution confocal laser endoscopic techniques (CLEs) during fluorescence-guided brain tumor surgery is a new tool for intraoperative tumor tissue visualization. To overcome observer dependent errors, we aimed to predict tumor type by applying a deep learning model to image data obtained by CLE.

Methods

Human brain tumor specimens from 25 patients with brain metastasis, glioblastoma, and meningioma were evaluated within this study. In addition to routine histopathological analysis, tissue samples were stained with fluorescein ex vivo and analyzed with CLE. We trained two convolutional neural networks and built a predictive level for the outputs.

Results

Multiple CLE images were obtained from each specimen with a total number of 13,972 fluorescein based images. Test accuracy of 90.9% was achieved after applying a two-class prediction for glioblastomas and brain metastases with an area under the curve (AUC) value of 0.92. For three class predictions, our model achieved a ratio of correct predicted label of 85.8% in the test set, which was confirmed with five-fold cross validation, without definition of confidence. Applying a confidence rate of 0.999 increased the prediction accuracy to 98.6% when images with substantial artifacts were excluded before the analysis. 36.3% of total images met the output criteria.

Conclusions

We trained a residual network model that allows automated, on site analysis of resected tumor specimens based on CLE image datasets. Further in vivo studies are required to assess the clinical benefit CLE can have.

SUBMITTER: Ziebart A

PROVIDER: S-EPMC8147727 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:Background: Previous studies showed that confocal laser endomicroscopy (CLE) images of brain tumors acquired by a first-generation (Gen1) CLE system using fluorescein sodium (FNa) contrast yielded a diagnostic accuracy similar to frozen surgical sections and histologic analysis. We investigated performance improvements of a second-generation (Gen2) CLE system designed specifically for neurosurgical use. Methods: Rodent glioma models were used for in vivo and rapid ex vivo CLE imaging. FNa and 5-aminolevulinic acid were used as contrast agents. Gen1 and Gen2 CLE images were compared to distinguish cytoarchitectural features of tumor mass and margin and surrounding and normal brain regions. We assessed imaging parameters (gain, laser power, brightness, scanning speed, imaging depth, and Z-stack [3D image acquisition]) and evaluated optimal values for better neurosurgical imaging performance with Gen2. Results: Efficacy of Gen1 and Gen2 was similar in identifying normal brain tissue, vasculature, and tumor cells in masses or at margins. Gen2 had smaller field of view, but higher image resolution, and sharper, clearer images. Other advantages of the Gen2 were auto-brightness correction, user interface, image metadata handling, and image transfer. CLE imaging with FNa allowed identification of nuclear and cytoplasmic contours in tumor cells. Injection of higher dosages of FNa (20 and 40 mg/kg vs. 0.1-8 mg/kg) resulted in better image clarity and structural identification. When used with 5-aminolevulinic acid, CLE was not able to detect individual glioma cells labeled with protoporphyrin IX, but overall fluorescence intensity was higher (p < 0.01) than in the normal hemisphere. Gen2 Z-stack imaging allowed a unique 3D image volume presentation through the focal depth. Conclusion: Compared with Gen1, advantages of Gen2 CLE included a more responsive and intuitive user interface, collection of metadata with each image, automatic Z-stack imaging, sharper images, and a sterile sheath. Shortcomings of Gen2 were a slightly slower maximal imaging speed and smaller field of view. Optimal Gen2 imaging parameters to visualize brain tumor cytoarchitecture with FNa as a fluorescent contrast were defined to aid further neurosurgical clinical in vivo and rapid ex vivo use. Further validation of the Gen2 CLE for microscopic visualization and diagnosis of brain tumors is ongoing.

Project description:BackgroundIntraoperative real-time confocal laser endomicroscopy (CLE) is an alternative modality for frozen tissue histology that enables visualization of the cytoarchitecture of living tissues with spatial resolution at the cellular level. We developed a new CLE with a "Lissajous scanning pattern" and conducted a study to identify its feasibility for fluorescence-guided brain tumor diagnosis.Materials and methodsConventional hematoxylin and eosin (H&E) histological images were compared with indocyanine green (ICG)-enhanced CLE images in two settings (1): experimental study with in vitro tumor cells and ex vivo glial tumors of mice, and (2) clinical evaluation with surgically resected human brain tumors. First, CLE images were obtained from cultured U87 and GL261 glioma cells. Then, U87 and GL261 tumor cells were implanted into the mouse brain, and H&E staining was compared with CLE images of normal and tumor tissues ex vivo. To determine the invasion of the normal brain, two types of patient-derived glioma cells (CSC2 and X01) were used for orthotopic intracranial tumor formation and compared using two methods (CLE vs. H&E staining). Second, in human brain tumors, tissue specimens from 69 patients were prospectively obtained after elective surgical resection and were also compared using two methods, namely, CLE and H&E staining. The comparison was performed by an experienced neuropathologist.ResultsWhen ICG was incubated in vitro, U87 and GL261 cell morphologies were well-defined in the CLE images and depended on dimethyl sulfoxide. Ex vivo examination of xenograft glioma tissues revealed dense and heterogeneous glioma cell cores and peritumoral necrosis using both methods. CLE images also detected invasive tumor cell clusters in the normal brain of the patient-derived glioma xenograft model, which corresponded to H&E staining. In human tissue specimens, CLE images effectively visualized the cytoarchitecture of the normal brain and tumors. In addition, pathognomonic microstructures according to tumor subtype were also clearly observed. Interestingly, in gliomas, the cellularity of the tumor and the density of streak-like patterns were significantly associated with tumor grade in the CLE images. Finally, panoramic view reconstruction was successfully conducted for visualizing a gross tissue morphology.ConclusionIn conclusion, the newly developed CLE with Lissajous laser scanning can be a helpful intraoperative device for the diagnosis, detection of tumor-free margins, and maximal safe resection of brain tumors.