Microscopy images are crucial for life science research, allowing detailed inspection and characterization of cellular and tissue-level structures and functions. However, microscopy data are unavoidably affected by image degradations, such as noise, blur, or others. Many such degradations also contribute to a loss of image contrast, which becomes especially pronounced in deeper regions of thick samples. Today, best performing methods to increase the quality of images are based on Deep Learning approaches, which typically require ground truth (GT) data during training. Our inability to counteract blurring and contrast loss when imaging deep into samples prevents the acquisition of such clean GT data. The fact that the forward process of blurring and contrast loss deep into tissue can be modeled, allowed us to propose a new method that can circumvent the problem of unobtainable GT data. To this end, we first synthetically degraded the quality of microscopy images even further by using an approximate forward model for deep tissue image degradations. Then we trained a neural network that learned the inverse of this degradation function from our generated pairs of raw and degraded images. We demonstrated that networks trained in this way can be used out-of-distribution (OOD) to improve the quality of less severely degraded images, e.g. the raw data imaged in a microscope. Since the absolute level of degradation in such microscopy images can be stronger than the additional degradation introduced by our forward model, we also explored the effect of iterative predictions. Here, we observed that in each iteration the measured image contrast kept improving while detailed structures in the images got increasingly removed. Therefore, dependent on the desired downstream analysis, a balance between contrast improvement and retention of image details has to be found.
The future of bioimage analysis is increasingly defined by the development and use of tools that rely on deep learning and artificial intelligence (AI). For this trend to continue in a way most useful for stimulating scientific progress, it will require our multidisciplinary community to work together, establish FAIR data sharing and deliver usable, reproducible analytical tools.
Light microscopy is routinely used to look at living cells and biological tissues at sub-cellular resolution. Components of the imaged cells can be highlighted using fluorescent labels, allowing biologists to investigate individual structures of interest. Given the complexity of biological processes, it is typically necessary to look at multiple structures simultaneously, typically via a temporal multiplexing scheme. Still, imaging more than 3 or 4 structures in this way is difficult for technical reasons and limits the rate of scientific progress in the life sciences. Hence, a computational method to split apart (decompose) superimposed biological structures acquired in a single image channel, i.e. without temporal multiplexing, would have tremendous impact. Here we present {\mu}Split, a dedicated approach for trained image decomposition. We find that best results using regular deep architectures is achieved when large image patches are used during training, making memory consumption the limiting factor to further improving performance. We therefore introduce lateral contextualization (LC), a memory efficient way to train deep networks that operate well on small input patches. In later layers, additional image context is fed at adequately lowered resolution. We integrate LC with Hierarchical Autoencoders and Hierarchical VAEs.For the latter, we also present a modified ELBO loss and show that it enables sound VAE training. We apply {\mu}Split to five decomposition tasks, one on a synthetic dataset, four others derived from two real microscopy datasets. LC consistently achieves SOTA results, while simultaneously requiring considerably less GPU memory than competing architectures not using LC. When introducing LC, results obtained with the above-mentioned vanilla architectures do on average improve by 2.36 dB (PSNR decibel), with individual improvements ranging from 0.9 to 3.4 dB.
In recent years, neural network based image denoising approaches have revolutionized the analysis of biomedical microscopy data. Self-supervised methods, such as Noise2Void (N2V), are applicable to virtually all noisy datasets, even without dedicated training data being available. Arguably, this facilitated the fast and widespread adoption of N2V throughout the life sciences. Unfortunately, the blind-spot training underlying N2V can lead to rather visible checkerboard artifacts, thereby reducing the quality of final predictions considerably. In this work, we present two modifications to the vanilla N2V setup that both help to reduce the unwanted artifacts considerably. Firstly, we propose a modified network architecture, i.e., using BlurPool instead of MaxPool layers throughout the used U-Net, rolling back the residual U-Net to a non-residual U-Net, and eliminating the skip connections at the uppermost U-Net level. Additionally, we propose new replacement strategies to determine the pixel intensity values that fill in the elected blind-spot pixels. We validate our modifications on a range of microscopy and natural image data. Based on added synthetic noise from multiple noise types and at varying amplitudes, we show that both proposed modifications push the current state-of-the-art for fully self-supervised image denoising.
Transformer architectures show spectacular performance on NLP tasks and have recently also been used for tasks such as image completion or image classification. Here we propose to use a sequential image representation, where each prefix of the complete sequence describes the whole image at reduced resolution. Using such Fourier Domain Encodings (FDEs), an auto-regressive image completion task is equivalent to predicting a higher resolution output given a low-resolution input. Additionally, we show that an encoder-decoder setup can be used to query arbitrary Fourier coefficients given a set of Fourier domain observations. We demonstrate the practicality of this approach in the context of computed tomography (CT) image reconstruction. In summary, we show that Fourier Image Transformer (FIT) can be used to solve relevant image analysis tasks in Fourier space, a domain inherently inaccessible to convolutional architectures.
Image denoising and artefact removal are complex inverse problems admitting many potential solutions. Variational Autoencoders (VAEs) can be used to learn a whole distribution of sensible solutions, from which one can sample efficiently. However, such a generative approach to image restoration is only studied in the context of pixel-wise noise removal (e.g. Poisson or Gaussian noise). While important, a plethora of application domains suffer from imaging artefacts (structured noises) that alter groups of pixels in correlated ways. In this work we show, for the first time, that generative diversity denoising (GDD) approaches can learn to remove structured noises without supervision. To this end, we investigate two existing GDD architectures, introduce a new one based on hierarchical VAEs, and compare their performances against a total of seven state-of-the-art baseline methods on five sources of structured noise (including tomography reconstruction artefacts and microscopy artefacts). We find that GDD methods outperform all unsupervised baselines and in many cases not lagging far behind supervised results (in some occasions even superseding them). In addition to structured noise removal, we also show that our new GDD method produces new state-of-the-art (SOTA) results on seven out of eight benchmark datasets for pixel-noise removal. Finally, we offer insights into the daunting question of how GDD methods distinguish structured noise, which we like to see removed, from image signals, which we want to see retained.
Automatic detection and segmentation of objects in microscopy images is important for many biological applications. In the domain of natural images, and in particular in the context of city street scenes, embedding-based instance segmentation leads to high-quality results. Inspired by this line of work, we introduce EmbedSeg, an end-to-end trainable deep learning method based on the work by Neven et al. While their approach embeds each pixel to the centroid of any given instance, in EmbedSeg, motivated by the complex shapes of biological objects, we propose to use the medoid instead. Additionally, we make use of a test-time augmentation scheme, and show that both suggested modifications improve the instance segmentation performance on biological microscopy datasets notably. We demonstrate that embedding-based instance segmentation achieves competitive results in comparison to state-of-the-art methods on diverse and biologically relevant microscopy datasets. Finally, we show that the overall pipeline has a small enough memory footprint to be used on virtually all CUDA enabled laptop hardware. Our open-source implementation is available at github.com/juglab/EmbedSeg.
Many microscopy applications are limited by the total amount of usable light and are consequently challenged by the resulting levels of noise in the acquired images. This problem is often addressed via (supervised) deep learning based denoising. Recently, by making assumptions about the noise statistics, self-supervised methods have emerged. Such methods are trained directly on the images that are to be denoised and do not require additional paired training data. While achieving remarkable results, self-supervised methods can produce high-frequency artifacts and achieve inferior results compared to supervised approaches. Here we present a novel way to improve the quality of self-supervised denoising. Considering that light microscopy images are usually diffraction-limited, we propose to include this knowledge in the denoising process. We assume the clean image to be the result of a convolution with a point spread function (PSF) and explicitly include this operation at the end of our neural network. As a consequence, we are able to eliminate high-frequency artifacts and achieve self-supervised results that are very close to the ones achieved with traditional supervised methods.
Deep Learning based methods have emerged as the indisputable leaders for virtually all image restoration tasks. Especially in the domain of microscopy images, various content-aware image restoration (CARE) approaches are now used to improve the interpretability of acquired data. But there are limitations to what can be restored in corrupted images, and any given method needs to make a sensible compromise between many possible clean signals when predicting a restored image. Here, we propose DivNoising -- a denoising approach based on fully-convolutional variational autoencoders, overcoming this problem by predicting a whole distribution of denoised images. Our method is unsupervised, requiring only noisy images and a description of the imaging noise, which can be measured or bootstrapped from noisy data. If desired, consensus predictions can be inferred from a set of DivNoising predictions, leading to competitive results with other unsupervised methods and, on occasion, even with the supervised state-of-the-art. DivNoising samples from the posterior enable a plethora of useful applications. We are (i) discussing how optical character recognition (OCR) applications could benefit from diverse predictions on ambiguous data, and (ii) show in detail how instance cell segmentation gains performance when using diverse DivNoising predictions.