Abstract
Image enhancement approaches often assume that noise is signal-independent, approximating the degradation model as zero-mean additive Gaussian. However, this assumption fails in biomedical imaging systems where sensor-based noise is proportional to signal strength, better represented as a Poisson process. We present Poisson2Sparse, a novel self-supervised learning method for single-image denoising where noise is approximated as a Poisson process, requiring no clean ground-truth data. By approximating traditional iterative optimization algorithms with a recurrent neural network (RNN) that enforces sparsity, our method suppresses spurious noise components through implicit regularization. Experiments on two bio-imaging datasets demonstrate that Poisson2Sparse outperforms state-of-the-art approaches in PSNR and SSIM, recovering subtle structural details that other methods often obscure.
Introduction
In the realm of computational imaging, the standard assumption of Additive White Gaussian Noise (AWGN) has driven the development of many successful denoising algorithms. However, in biomedical contexts—such as fluorescence microscopy or X-ray imaging—the physics of photon counting dictates that noise is signal-dependent. In these scenarios, the noise follows a Poisson distribution, where the variance is tied directly to the signal intensity. Brighter regions are inherently noisier, a characteristic that standard Gaussian-based models fail to capture.
Furthermore, supervised deep learning models typically require large datasets of noisy/clean image pairs. In clinical or biological research, obtaining "clean" ground truth is often impossible due to the dynamic nature of samples (e.g., a beating heart) or the risk of phototoxicity from long exposure times. Poisson2Sparse addresses these challenges by utilizing a self-supervised framework that learns to denoise using only the single noisy image provided at test time.
Methodology
Poisson2Sparse bridges the gap between classical signal processing and modern deep learning by "unrolling" traditional optimization into a neural architecture.
1. Unrolled Optimization and Sparsity
The core of the method is based on the Iterative Shrinkage Thresholding Algorithm (ISTA), a classical approach for solving sparse reconstruction problems. The authors approximate the steps of ISTA using a Recurrent Neural Network (RNN) structure. Instead of using fixed mathematical operators, the network employs a learnable Encoder-Decoder structure to identify a "sparse representation" of the image. Because biological structures are repetitive and noise is random (and thus non-sparse), forcing the network to find a sparse representation naturally filters out the noise as spurious components.
2. Convolutional Sparse Coding (CSC)
To ensure the model is agnostic to input dimensions and can capture local features effectively, the framework utilizes Convolutional Sparse Coding. By learning a "dictionary" of image patches via convolutions, the network recognizes recurring structural motifs—such as cell membranes or vessel walls—across the entire image, reinforcing the underlying signal over the random noise.
3. Self-Supervised Training via Neighbor Down-sampling
To facilitate training without ground truth, the authors employ Random Neighbor Down-sampling (RND). The single noisy input is split into two sub-sampled versions. Since adjacent pixels in a physical image are highly correlated while their noise realizations are independent, the network can be trained to predict the underlying signal of one sub-sample from the other.
4. Physics-Informed Loss Function
Rather than using the standard Mean Squared Error (MSE) loss, which is mathematically derived from Gaussian assumptions, Poisson2Sparse utilizes a loss function based on the maximum log-likelihood of a Poisson process. This ensures the optimization is physically consistent with the sensor data.
Experimental Results
The authors evaluated Poisson2Sparse against two primary datasets: the Fluorescence Microscopy Denoising (FMD) dataset and the PINCAT simulated cardiac perfusion MRI dataset.
Quantitative Performance
On the PINCAT dataset, Poisson2Sparse demonstrated significant gains, outperforming state-of-the-art methods like BM3D (traditional) and DIP/Self2Self (self-supervised) by approximately 2 dB in Peak Signal-to-Noise Ratio (PSNR). While some competitors showed instability—occasionally producing blank or corrupted images—Poisson2Sparse remained robust across varying noise intensities.
Qualitative Detail Recovery
Beyond metrics, the qualitative results highlight the method's ability to preserve "subtle details." In microscopy and cardiac imaging, preserving the sharpness of thin lines (like the rib cage or fine cellular filaments) is critical. While other self-supervised methods tended to over-smooth these features into blurry blobs, Poisson2Sparse maintained structural integrity and texture.
Discussion and Limitations
While Poisson2Sparse represents a significant step forward for single-image denoising, it faces certain practical constraints. Because the model undergoes "internal learning" (training specifically on the test image), it requires approximately 7 minutes of GPU time for a 128x128 image. This computational overhead makes it less suitable for real-time applications compared to pre-trained supervised models.
Additionally, real-world sensors often exhibit a mixture of Poisson (shot) and Gaussian (read) noise. In the FMD dataset, where noise is not "purely" Poisson, the method's SSIM scores were slightly lower than some competitors, despite superior visual clarity. Future work aims to extend this framework to handle complex Poisson-Gaussian mixtures, further broadening its utility for commercial imaging hardware.
Conclusion
Poisson2Sparse demonstrates that the combination of classical sparsity priors and self-supervised deep learning can effectively tackle the challenges of biomedical image denoising. By tailoring the architecture and loss function to the physics of Poisson noise, the method provides a robust, ground-truth-free solution for recovering high-fidelity biological details from noisy, single-frame captures.
Summary updated