Academic Background

Voltage imaging is a powerful technique for studying neuronal activity, but its effectiveness is often constrained by low signal-to-noise ratios (SNR). Traditional denoising methods, such as matrix factorization, impose rigid assumptions about noise and signal structures, while existing deep learning approaches fail to fully capture the rapid dynamics and complex dependencies inherent in voltage imaging data. To address these issues, this paper introduces CellMincer, a novel self-supervised deep learning method specifically developed for denoising voltage imaging datasets. CellMincer operates by masking and predicting sparse pixel sets across short temporal windows and conditions the denoiser on precomputed spatiotemporal auto-correlations to effectively model long-range dependencies, thereby significantly improving denoising performance.

Voltage imaging utilizes fluorescent reporters, such as small-molecule dyes or genetically encoded proteins, to measure the membrane potential of electrically active cells. Compared to traditional patch-clamp electrophysiology (EP), voltage imaging offers higher throughput and is less invasive. However, due to factors such as dye quantum yield, short exposure times (<2ms), and constraints on excitation intensity, the achievable SNR remains limited, which can obscure small-magnitude electrical events of interest, such as subthreshold post-synaptic potentials, and hinder the understanding of neuronal circuit formation and synaptic plasticity.

Paper Source

This paper is co-authored by Brice Wang, Tianle Ma, Theresa Chen, Trinh Nguyen, Ethan Crouse, Stephen J. Fleming, Alison S. Walker, Vera Valakh, Ralda Nehme, Evan W. Miller, Samouil L. Farhi, and Mehrtash Babadi. The authors are affiliated with institutions such as the Broad Institute of MIT and Harvard, Oakland University, and UC Berkeley. The paper was published in npj Imaging in 2024.

Research Process

1. Data Preprocessing and Global Feature Extraction

The CellMincer denoising pipeline involves three stages: data preprocessing, self-supervised pretraining, and denoising inference. In the preprocessing stage, researchers represent voltage imaging data as a three-dimensional tensor (time × width × height) and fit a low-order polynomial to each pixel’s time series, decomposing it into a smooth trend and a detrended residual tensor. The trend tensor primarily represents background fluorescence, while the detrended residual tensor represents noisy measurements of electrical activity. This step removes background noise unrelated to the signal, thereby improving model performance.

2. Self-Supervised Pretraining

In the self-supervised pretraining stage, CellMincer trains by masking and predicting sparse pixel sets. Specifically, researchers randomly select pixels in a frame and replace them with Gaussian noise, then train the neural network to predict the values of these masked pixels from the remaining pixels. This training strategy allows the model to denoise without requiring clean target data, circumventing the need for traditional supervised learning methods that rely on clean data.

3. Denoising Inference

During the inference stage, CellMincer inputs the detrended movie data into the neural network and denoises the middle frame of each sliding window. To avoid truncated results, researchers apply appropriate spatial padding during training and inference and add τ copies of the first and last frames to the beginning and end of the denoised movie, respectively.

4. Physics-Based Simulation Framework

To optimize CellMincer’s architecture and hyperparameters, researchers developed OptoSynth, a physics-based simulation framework for generating highly realistic synthetic voltage imaging datasets. OptoSynth simulates noiseless voltage imaging readouts by modeling neuron morphology reconstructions and patch-clamp electrophysiology measurements, then adds Poisson shot noise and Gaussian sensor thermal noise to generate realistic noisy data. These synthetic datasets are used for rigorous hyperparameter optimization and ablation studies, highlighting the critical role of precomputed spatiotemporal auto-correlations in denoising.

Key Results

1. Denoising Performance

Comprehensive benchmarking on both simulated and real datasets demonstrates that CellMincer achieves state-of-the-art performance in terms of SNR gain, high-frequency noise reduction, subthreshold event detection, and EP signal recovery. Compared to existing methods, CellMincer improves SNR gain by 0.5–2.9 dB and reduces SNR variability by 17–55%. Notably, CellMincer achieves a 14 dB reduction in high-frequency noise (>100Hz), outperforming the next best methods by an additional 3–10.5 dB.

2. Subthreshold Event Detection

On real voltage imaging data, CellMincer significantly improves the detection accuracy of subthreshold events. Compared to benchmarked methods, CellMincer increases the F1-score for detecting subthreshold events by 2–6 percentage points across voltage magnitudes in the 0.5–10 mV range. Additionally, CellMincer improves the cross-correlation between low-noise EP recordings and voltage imaging by 8%.

3. Neuron Segmentation and Functional Phenotype Identification

The incorporation of CellMincer into standard workflows significantly improves neuron segmentation, peak detection, and functional phenotype identification. In voltage imaging of chronically tetrodotoxin (TTX)-treated and unperturbed hPSC-derived neurons, CellMincer denoising enables the reliable identification and segmentation of nearly twice as many neurons as in the raw data, ultimately enhancing the statistical separation between the two functional phenotypes.

Conclusion

CellMincer is a self-supervised deep learning method specifically designed for denoising voltage imaging datasets. By masking and predicting sparse pixel sets and conditioning the denoiser on precomputed spatiotemporal auto-correlations, CellMincer significantly improves denoising performance. Its state-of-the-art results on both simulated and real datasets, particularly in SNR gain, high-frequency noise reduction, and subthreshold event detection, make it a powerful tool for advancing the study of neuronal circuits and synaptic plasticity.

Research Highlights

1. Innovative Method: CellMincer leverages self-supervised learning and precomputed spatiotemporal auto-correlations to significantly enhance the denoising of voltage imaging data.
2. Broad Applicability: CellMincer outperforms existing methods on both simulated and real datasets, particularly in high-frequency noise reduction and subthreshold event detection.
3. Practical Value: The integration of CellMincer improves neuron segmentation, peak detection, and functional phenotype identification, providing new tools for neuronal circuit research and synaptic plasticity analysis.

Additional Valuable Information

CellMincer’s code release emphasizes usability and ease of deployment, offering various diagnostic feedback mechanisms and a stable Docker image for public use. Additionally, researchers have made pre-trained CellMincer models available, further reducing the computational cost of adopting this method.