Fluorescent functional imaging is crucial in neuroscience, aiming for wider FOV, deeper penetration, and faster speeds while improving SNR. Current techniques like point-scanning or spinning disk confocal systems compromise between SNR and parameters such as imaging speed, FOV, resolution, and recording length. While advancements in indicators and microscopy improve spatiotemporal resolution and FOV, trade-offs remain, particularly in organisms like C. elegans where low exposure times limit SNR to avoid motion artifacts. Increasing laser power improves SNR but causes photobleaching and phototoxicity. Deep learning has shown promise in enhancing microscopy, but existing methods require specialized expertise or equipment. Supervised deep learning for image denoising offers an alternative, potentially achieving higher accuracy and generalizability than unsupervised methods. However, supervised methods haven't been widely used for video denoising and calcium trace extraction due to challenges in data acquisition and the uncertainty of preserving temporal features in independently denoised images. This study explores a supervised deep-denoising approach to overcome these limitations and improve calcium trace extraction from noisy videos.

Recent advancements in deep learning have significantly improved the capabilities of microscopic techniques, enabling researchers to overcome tradeoffs between imaging speed and SNR. However, these techniques often require specialized knowledge or advanced microscopy setups. For example, some methods require detailed characterization of the microscopy system or specialized light-field microscopy. Deep-learning-based image denoising has emerged as a promising alternative. Both supervised and unsupervised methods have been developed. Unsupervised methods use the data itself for training, eliminating the need for separate training data. However, these often require large datasets and pre-registered videos, limiting their applicability. Supervised methods, while potentially more accurate and generalizable, have seen limited application in video denoising and calcium trace extraction for several reasons, including the difficulty of collecting paired low and high SNR video data.

To address these challenges, the authors developed Neuro-Imaging Denoising via Deep Learning (NIDDL), a convolutional neural network (CNN) pipeline. NIDDL uses independently acquired image pairs (noisy and high SNR) for training, improving generalizability and ease of use. The pipeline takes noisy and high SNR image stacks as input, trains efficient denoising CNNs, applies the trained networks to denoise video data independently frame-by-frame, and finally, extracts high-quality calcium traces using a standard pipeline (cell segmentation, tracking, signal extraction). Network optimization involved testing various architectures (UNet, Hourglass), hyperparameters (kernel size, channel depth, residual connections), loss functions (L2, L1), and training modes (2D, 2.5D, 3D). The optimal models minimized the number of parameters and memory footprint by fixing channel depth across layers. This resulted in deeper networks with residual connections, offering significant improvements in model size, training time, and inference speed compared to existing methods like CARE and RCAN. The optimal models were trained with a significantly smaller number of image pairs (500-600) compared to previous methods. The authors also tested the networks' robustness against different noise levels using synthetic data generated with Poisson shot noise and Gaussian readout noise and evaluated performance across various metrics (RMSE, PSNR, SSIM). Three types of calcium imaging data were used for evaluation: high-magnification whole-brain recordings, low-magnification large FOV recordings, and high-magnification neurite recordings. Calcium trace extraction involved denoising videos, segmenting nuclei (for whole-brain data), tracking cells, and extracting traces (single-pixel or ROI averaged). For freely moving animals, manual tracking was used. Neurite segmentation utilized image sharpening, adaptive thresholding, erosion, and removal of small structures. The performance of NIDDL was compared against several traditional and deep-learning-based methods (Median filtering, Gaussian filtering, NLM, BM3D, CARE, RCAN).

NIDDL achieved significant improvements in denoising accuracy across all three applications compared to traditional and other deep learning methods. The optimized models are 20-30 times smaller, with 3-4 times faster inference speeds than CARE and RCAN. Training required only 500-600 image pairs, far fewer than previous methods. NIDDL effectively recovered fine structures like nuclei and neurites, enhancing downstream analysis. The method generalized well across different strains, experimental sessions (with the exception of laser power), and noise levels. Denoising with NIDDL resulted in high-quality calcium traces with lower mean absolute error and higher correlation to ground truth traces. It also successfully preserved the correlational structure among neuronal activities, crucial for downstream analyses like PCA. For large FOV imaging, NIDDL enabled the detection of cells even with low spatial resolution and low laser power, avoiding photobleaching. In freely moving animals, NIDDL effectively denoised traces, revealing coordinated neuronal activity correlated with body curvature. For neurite imaging, NIDDL improved neurite segmentation, allowing quantitative characterization of neurite morphology and generalizing well across different strains with varying neurite structures.

NIDDL addresses critical limitations of existing deep learning methods for neuro-imaging denoising. Its ability to train with temporally independent data simplifies data acquisition and eliminates the need for ultrafast imaging rates or complex pre-registration steps, making it more accessible. The small training dataset requirement reduces the barrier to entry for many researchers. The high generalizability across diverse conditions and strains makes the method robust and widely applicable. The improved speed and efficiency enable real-time applications and online feedback manipulations. The successful application across various imaging scenarios (whole-brain imaging, large FOV, and neurite imaging) demonstrates NIDDL's versatility and broad impact on neuroscience research.

NIDDL provides a significant advancement in neuro-imaging denoising by offering an efficient, accurate, and generalizable deep learning framework. Its ease of training, reduced data requirements, and fast inference speed make it a valuable tool for researchers working with various imaging modalities and model organisms. Future research could explore the application of NIDDL to other imaging techniques and expand its capabilities for more complex downstream analyses.

While NIDDL shows excellent performance, there are some limitations. The model's performance is sensitive to the laser power used during image acquisition. Additionally, the generalizability across laser powers might need further investigation for applications with a wide range of SNR levels. Although the authors used a semi-synthetic video dataset, testing on more diverse and complex real-world datasets would further strengthen the robustness claims.