DeepContrast: Deep Tissue Contrast Enhancement using Synthetic Data Degradations and OOD Model Predictions

TL;DR

This paper tackles the challenge of improving contrast in deep-tissue microscopy where ground-truth data are unavailable. It introduces DeepContrast, which synthetically degrades raw images via a forward model $d(x) = \alpha \cdot x + (1-\alpha) \cdot n(b(x))$ to create paired training data for a bias-free U-Net that inverts the degradation, enabling OOD enhancement of raw images. Across liver tissue datasets, DeepContrast outperforms CLAHE, Huygens deconvolution, and FCE-Net in contrast metrics and downstream segmentation, with iterative applications offering greater contrast while preserving more structures than baselines. The study provides a practical, iterative framework for enhancing deep-tissue image quality with quantified trade-offs between contrast and detail, suitable for downstream quantitative analyses in biological imaging.

Abstract

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.

Figures (7)

• Figure 1: Proposed scheme to improve deep tissue contrast. (1) Pairs of data for supervised training are generated by degrading raw microscopy images using a suitable degradation function $d(x)$ composed of a blurring and a noising step. (2) During supervised network training, synthetically degraded images are used as inputs and the original images as targets. (3) During inference, we feed the original raw microscopy images once or iteratively into the trained network (see Section \ref{['sec:methods_iterative_pred']}).
• Figure 2: Qualitative results. Images of liver tissue sections stained with Phalloidin as a proxy for cell borders, used to compare our results (DeepContrast Model A) to several baseline methods. Rows show image planes at different depths in the liver tissue. Columns show the raw input, results obtained with CLAHE zuiderveld_contrast_1994, Huygens deconvolution (see Section \ref{['sec:experiments_baselines']}), best FCE-Net zhang_fce-net_2022 results ($3\times$), and our best DeepContrast results ($3\times$), respectively. The three rightmost columns shows the inset areas marked by dashed boxes and line plots of raw intensities, the FCE-Net, and DeepContrast (along the green line in the respective images). Scale bars: $20 \mu m$ in full size images, $10 \mu m$ in insets.
• Figure 3: Qualitative results of iterative OOD model application. Contrast of the image data of Figure \ref{['fig:results_qualitative']} iteratively enhanced using a trained DeepContrast network (Model A). Rows show, analogous to Figure \ref{['fig:results_qualitative']}, image planes at different depth into the imaged tissue. Columns show the raw input data, and the results of applying DeepContrast a single time, two, three (same as in Figure \ref{['fig:results_qualitative']}), and six consecutive times. The three rightmost columns shows the inset areas marked by dashed boxes and line plots along the green lines in the raw data, and along the $1\times$, $3\times$, and $6\times$ enhanced outputs. Note that, while contrast is continuously enhanced, too many iterative applications cause a notable loss of image details. Scale bars: 20 $\mu m$ in full size images, 10 $\mu m$ in insets.
• Figure 4: Quantitative results. Contrast quantification using the Percentile Contrast Index (see Section \ref{['sec:contrast_quantification']}) and the Wavelet Contrast Index albright_paintings_2023 (higher values are better) represented as average and $95\%$ Confidence Intervals at each depth ($N=18$). Dashed vertical grey lines depict depths shown in Figure \ref{['fig:results_qualitative']}. Multiple iterations of FCE-Net zhang_fce-net_2022 and our DeepContrast (Model A) approach show image contrast is further improved when these networks are iteratively applied.
• Figure 5: Qualitative and quantitative results of segmentation masks created at multiple iterations of contrast enhancement. Left side column shows Raw input and segmentation mask (Section \ref{['sec:downstream_analysis']}). Each further column shows different iterations of contrast enhancement and corresponding segmentation of cell borders. Top row shows inference results with DeepContrast and bottom row shows inference results with FCE-Net. Yellow arrow heads in images highlight lost or degraded structures when comparing DeepContrast and FCE-Net. Violin plots shows distribution of IoU values between the different contrast enhancement methods and raw segmentation masks used as reference at multiple iterations ($N=2682$), showing faster decrease of IoU values and more abundant mistakes in segmentation with FCE-Net.