Deep Joint Denoising and Detection for Enhanced Intracellular Particle Analysis

TL;DR

This work tackles reliable intracellular particle analysis in noisy time-lapse fluorescence images. It introduces DENODET, a one-encoder-two-parallel-decoders network that jointly denoises and detects particles, leveraging a shared encoder and a DSNT-based localization pathway. The training uses a composite loss, combining $L_{DET}$ and $L_{DENO}$ in the form $\mathcal{L} = L_{DET} + \gamma L_{DENO}$ with $\gamma \approx 0.5$, and employs a Dice-like term plus balanced cross-entropy for denoising and Euclidean plus Jensen–Shannon divergence for detection. Empirical results on synthetic and real 2D+t data show that DENODET achieves superior detection (F1) and localization (RMSE) compared to non-denoising baselines, cascaded denoising–detection, and competing deep detectors, demonstrating the benefit of integrated denoising for particle analysis. The method holds promise for improving particle tracking performance in fluorescence microscopy and can be extended to 3D+t and other high-level detection tasks.

Abstract

Reliable analysis of intracellular dynamic processes in time-lapse fluorescence microscopy images requires complete and accurate tracking of all small particles in all time frames of the image sequences. A fundamental first step towards this goal is particle detection. Given the small size of the particles, their detection is greatly affected by image noise. Recent studies have shown that applying image denoising as a preprocessing step indeed improves particle detection and their subsequent tracking. Deep learning based particle detection methods have shown superior results compared to traditional detection methods. However, they do not explicitly aim to remove noise from the images to facilitate detection. Thus we hypothesize that their performance could be further improved. In this paper, we propose a new deep neural network, called DENODET (denoising-detection network), which performs image denoising and particle detection simultaneously. We show that integrative denoising and detection yields more accurate detection results. Our method achieves superior results compared to state-of-the-art particle detection methods on the particle tracking challenge dataset and our own real fluorescence microscopy image data.

Figures (4)

• Figure 1: Conventional particle detection approaches and our proposed approach. a) Separate denoising and detection networks trained separately. The denoising network takes a noisy input image and outputs an enhanced image. The detection network takes the noisy input image and outputs the particle positions. b) Cascaded approach where the input image is passed to a denoising network and then the enhanced output image is passed to a detection network to extract particle positions. c) Our proposed joint framework for image denoising and particle detection. Our network has an encoder-two-parallel-decoders structure. One of the decoders generates a denoised image while the other detects particles in the noisy input image with skip connections between the encoder and the two parallel decoders.
• Figure 2: Architecture of the proposed network. a) Overview of our one-encoder-dual-decoder network design for performing joint image denoising and particle detection. The solid blue lines indicate the denoising decoder path, the solid green lines indicate the detection decoder path, and the dashed red lines indicate skip connections from the encoder to the two decoders, and from the denoising decoder to the detection decoder in each level. b) Architecture of the residual convolutional block. c) Architecture of the upsampling block.
• Figure 3: Example detection results of DENODET on synthetic images of vesicles, receptors, and microtubule plus-ends, at SNR levels 1, 2, 3, 4, and 5. The red circles indicate the predicted positions and the green circles indicate the reference positions.
• Figure 4: Example results of applying detection methods SOS, Detnet, and DENODET to real fluorescence microscopy image data showing different markers Rab5, Rab6, Rab11, and EB3. Reference particle positions manually annotated by an expert biologist are shown in the first row for comparison. Red arrows point at prominent false positive and false negative detections compared to the expert manual annotations.