Learned, uncertainty-driven adaptive acquisition for photon-efficient scanning microscopy

TL;DR

A method to simultaneously denoise and predict pixel-wise uncertainty for scanning microscopy systems, improving algorithm trustworthiness and providing statistical guarantees for deep learning predictions is proposed and the first to demonstrate distribution-free uncertainty quantification for a denoising task with real experimental data and the first to propose adaptive acquisition based on reconstruction uncertainty.

Abstract

Scanning microscopy systems, such as confocal and multiphoton microscopy, are powerful imaging tools for probing deep into biological tissue. However, scanning systems have an inherent trade-off between acquisition time, field of view, phototoxicity, and image quality, often resulting in noisy measurements when fast, large field of view, and/or gentle imaging is needed. Deep learning could be used to denoise noisy microscopy measurements, but these algorithms can be prone to hallucination, which can be disastrous for medical and scientific applications. We propose a method to simultaneously denoise and predict pixel-wise uncertainty for scanning microscopy systems, improving algorithm trustworthiness and providing statistical guarantees for deep learning predictions. Furthermore, we propose to leverage this learned, pixel-wise uncertainty to drive an adaptive acquisition technique that rescans only the most uncertain regions of a sample, saving time and reducing the total light dose to the sample. We demonstrate our method on experimental confocal and multiphoton microscopy systems, showing that our uncertainty maps can pinpoint hallucinations in the deep learned predictions. Finally, with our adaptive acquisition technique, we demonstrate up to 16X reduction in acquisition time and total light dose while successfully recovering fine features in the sample and reducing hallucinations. We are the first to demonstrate distribution-free uncertainty quantification for a denoising task with real experimental data and the first to propose adaptive acquisition based on reconstruction uncertainty.

Figures (8)

• Figure 1: Uncertainty-based Adaptive Imaging (a): A noisy measurement is acquired with a scanning microscopy system and passed into a deep learning model that predicts a denoised image and its associated pixel-wise uncertainty. Subsequently, the top N uncertain pixels are selected for a rescan, obtaining more measurements at only the uncertain regions. As more adaptive measurements are taken, the deep learning model predicts a denoised image with lower uncertainty. Scan duration and power are minimized, limiting sample damage while maintaining high confidence in the model prediction. Rescanning Process (b): Given a pixel-wise uncertainty prediction, regions with high uncertainty can be selected for rescanning. Only this patch of pixels will be rescanned in the sample, and this patch, superimposed with the original, becomes an additional channel that is fed into the model.
• Figure 2: Learned Uncertainty Quantification: To leverage distribution-free uncertainty quantification with a deep network, three modifications are needed: a last layer that predicts a lower and upper bound, a modified loss, and a post-training calibration step. Standard deep denoising steps are highlighted in blue, while our modifications are highlighted in green. (a) During training, the modified network returns three channels: the lower uncertainty quantile, the denoised prediction, and the upper uncertainty quantile. A quantile loss function defines the loss for the upper and lower quantiles, encouraging underestimates and overestimates. (b) After training, a calibration step is needed to adjust the upper and lower bounds and provide statistical guarantees for this predicted interval. (c) At test time, a single forward pass determines the uncertainty at each step, which is the difference between the upper and lower quantile predictions.
• Figure 3: Quantile regression. An asymmetric loss function is used to encourage the deep network to predict the upper and lower-bound images. As opposed to a symmetric loss like MSE, which penalizes estimates above and below the mean value equally, quantile loss is asymmetric. When $\alpha$ is small (bottom), predictions higher than the mean (over-estimates) are more heavily penalized than predictions lower than the mean (under-estimates), encouraging the network to predict an image that is below the mean (lower bound image). When $\alpha$ is large (top), under-estimates are more heavily penalized than under-estimates, encouraging the network to predict an image that is above the mean (upper bound image).
• Figure 4: Multi-image FMD Denoising Results. We compare single and multi-image denoising on a representative two-photon sample from the FMD dataset. As the number of measurements increases from 1 to 6 and 10, sharper features emerge in the denoised images (top row) and the predicted uncertainty (2nd from top row) decreases. Quantitatively, the MSE and SSIM improve over iterations, and the average uncertainty decreases (bottom row). Zooming in on a small region within the sample (inset), we can see that single-image denoising produces a hallucination (red box): a horizontal feature that is not present in the ground truth image. This hallucination has high uncertainty. After multi-image denoising, the hallucination goes away as the predicted image converges to the ground truth feature, and concurrently the uncertainty in this region decreases.
• Figure 5: Multi-image MPM Denoising Results. Next, we compare single and multi-image denoising on a 2PAF sample from our custom MPM dataset, using a network that has never seen MPM data. As before, multi-image denoising results in sharper features and a lower uncertainty (top two rows), which can be quantified by the MSE, SSIM, and average uncertainty (bottom). Single-image denoising results in a hallucination (red box) which is not present in the ground truth.