Background: Confocal laser endomicroscopy (CLE) allows imaging of tissue autofluorescence (AF) in situ and is promising for brain tumor margin detection and diagnosis. However, comprehensive analysis of CLE images beyond visual inspection is restricted by their high noise level. Because image acquisition parameters cannot be accordingly optimized in medical devices, image preprocessing constitutes an option for noise removal. Here, we utilized an AI-based algorithm trained on astronomy images and investigated its capacity to denoise AF images of brain tumors. Methods: Tissue samples of human brain tumors and non-neoplastic brain were investigated. Images of tissue AF (518–573 nm) were acquired on bulk tissue samples using a clinical CLE system with 488 nm excitation. Multiphoton microscopy was employed to investigate tissue sections and image averaging was used to modify the noise level in AF images (500–550 nm). Signal to noise ratio (SNR) of AF images were compared before and after denoising with NoiseXTerminator software. Results: Denoising improved SNR of multiphoton and CLE images and intensity plots indicated a profound improvement in the visibility of structures. Comparison of denoised high noise (4× and 8× averaging) with low noise (16× averaging) multiphoton images confirmed that image structures were correctly extracted. Upon the introduction of synthetic noise, Gaussian noise was eliminated, however limitations were identified in the ability to remove bright pixels induced by salt-and-pepper noise. Conclusions: Efficient denoising of CLE images of tissue AF constitutes the basis for mathematical analysis and will allow the development of objective clinical protocols for intraoperative decision support in neurosurgery.