Objective: With the rapid growth of high-speed deep-tissue imaging in biomedical research, there is an urgent need to develop a robust and effective denoising method to retain morphological features for further texture analysis and segmentation. Conventional denoising lters and models can easily suppress the perturbative noise in high-contrast images; however, for low photon budget multiphoton images, a high detector gain will not only boost the signals but also bring signicant background noise. In such a stochastic resonance imaging regime, subthreshold signals may be detectable with the help of noise, meaning that a denoising lter capable of removing noise without sacricing important cellular features, such as cell boundaries, is desirable. Method: We propose a convolutional neural network-based denoising autoencoder methoda fully convolutional deep denoising autoencoder (DDAE)to improve the quality of three-photon uorescence (3PF) and third-harmonic generation (THG) microscopy images.

Results: The average of 200 acquired images of a given location served as the low-noise answer for the DDAE training. Compared with other conventional denoising methods, our DDAE model shows a better signal-to-noise ratio (28.86 and 21.66 for 3PF and THG, respectively), structural similarity (0.89 and 0.70 for 3PF and THG, respectively), and preservation of the nuclear or cellular boundaries (F1-score of 0.662 and 0.736 for 3PF and THG, respectively). It shows that DDAE is a better trade-off approach between structural similarity and preserving signal regions. Conclusions: The results of this study validate the effectiveness of the DDAE system in boundary-preserved image denoising. Clinical Impact: The proposed deep denoising system can enhance the quality of microscopic images and effectively support clinical evaluation and assessment.