Boundary-Preserved Deep Denoising of Stochastic Resonance Enhanced Multiphoton Images

Abstract

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 filters 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 significant background noise. In such a stochastic resonance imaging regime, subthreshold signals may be detectable with the help of noise, meaning that a denoising filter capable of removing noise without sacrificing important cellular features, such as cell boundaries, is desirable.

Method: We propose a convolutional neural network-based denoising autoencoder method - a fully convolutional deep denoising autoencoder (DDAE) - to improve the quality of three-photon fluorescence (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.

Keywords: Third harmonic generation; deep denoising autoencoder; three-photon fluorescence.

FIGURE 1.

The procedures of image processing for analyzing the preservation of nuclear and cellular boundaries.

FIGURE 2.

(a) The boundaries (yellow contours) of the binarized 3PF answer image bImag can precisely outline the nuclear boundaries. (b) The superposition image of the true positive (blue color), the false positive (magenta color), and the false negative (green color) parts of denoised THG images. Fields of view: (a) ; (b) .

FIGURE 3.

(a) Noisy inputs (upper rows) and low-noise answers (bottom rows) of three testing 3PF images of Hoechst blue labeled RAW cells. (b) The low-noise answer image, (c) noisy input, and (d) DDAE processed one. Processed by DDAE model, the noise was suppressed, the contrast was enhanced, and the nuclear boundary was well-preserved (yellow dashed closure). Fields of views: (a) ; (b–d) .

FIGURE 4.

Denoising results of 3PF images with DDAE, Gaussian filter, median filter, and BM3D algorithms.

FIGURE 5.

Noisy inputs (upper rows) and low-noise answers (bottom rows) of three testing THG images of RAW cells. Fields of views: .

FIGURE 6.

Denoising results of THG images with DDAE, Gaussian filter, median filter, and BM3D algorithms.

FIGURE 7.

Nuclear region analysis of 3PF images denoised with DDAE, Gaussian filter ( ), median filter ( ), and BM3D ( ) algorithms. Blue: true positive, Magenta: false positive, Green: false negative.

FIGURE 8.

Cellular region analysis of THG images denoised with DDAE, Gaussian filter ( ), median filter ( ), and BM3D ( ) algorithms. Blue: true positive, Magenta: false positive, Green: false negative.

FIGURE 9.

Training and validation loss across different epoch iteration during the DDAE training of 3PF images.

FIGURE 10.

Training and validation loss across different epoch iteration during the DDAE training of THG images.