Uncertainty-Guided Selective Adaptation Enables Cross-Platform Predictive Fluorescence Microscopy

TL;DR

The paper addresses the brittleness of deep learning models in fluorescence microscopy when deployed on new instruments and settings. It proposes SIT-ADDA-Auto, a self-configuring, uncertainty-guided extension of ADDA that selectively freezes deeper layers and adapts only early layers, with adaptation depth chosen from an ensemble-based uncertainty measure applied to unlabeled targets. Across cross-modality, cross-platform, and exposure/illumination/magnification shifts, SIT-ADDA improves reconstruction quality and downstream segmentation relative to full-encoder ADDA and non-adversarial baselines, while preserving semantic content. The work offers a practical, label-free strategy for cross-platform virtual staining and domain adaptation in microscopy, with potential applicability to other imaging domains and settings where labeled target data are scarce.

Abstract

Deep learning is transforming microscopy, yet models often fail when applied to images from new instruments or acquisition settings. Conventional adversarial domain adaptation (ADDA) retrains entire networks, often disrupting learned semantic representations. Here, we overturn this paradigm by showing that adapting only the earliest convolutional layers, while freezing deeper layers, yields reliable transfer. Building on this principle, we introduce Subnetwork Image Translation ADDA with automatic depth selection (SIT-ADDA-Auto), a self-configuring framework that integrates shallow-layer adversarial alignment with predictive uncertainty to automatically select adaptation depth without target labels. We demonstrate robustness via multi-metric evaluation, blinded expert assessment, and uncertainty-depth ablations. Across exposure and illumination shifts, cross-instrument transfer, and multiple stains, SIT-ADDA improves reconstruction and downstream segmentation over full-encoder adaptation and non-adversarial baselines, with reduced drift of semantic features. Our results provide a design rule for label-free adaptation in microscopy and a recipe for field settings; the code is publicly available.

Figures (6)

• Figure 1: Adversarial Discriminative Domain Adaptation (ADDA) enables robust cross-domain image translation in microscopy. (A) Schematic of the ADDA workflow. Stage 1 (Parental training): train an encoder and predictor (e.g., U-Net) on labeled source images. Stage 2 (Adversarial learning): adapt a target encoder to match the source using unlabeled target data and a discriminator. Stage 3: perform inference on target images with the adapted encoder and fixed predictor. Our variant, SIT-ADDA, selectively fine-tunes a subset of layers to improve robustness across microscopy platforms. (B) This framework generalizes across diverse imaging domain shifts, including differences in magnification, imaging modality, illumination conditions, and exposure settings. (C) Example application to cross-modality nuclear fluorescence prediction. A U-Net trained on phase-contrast data fails to generalize to brightfield inputs (middle), introducing artifacts (white arrows) and missing nuclei. In contrast, ADDA adaptation (right) preserves nuclear morphology and inter-nuclear variability in the target domain. (D) Quantitative evaluation using pixel-wise Pearson correlation shows that direct transfer from phase contrast to brightfield causes accuracy to collapse from $\sim$0.98 to $\sim$0.16, while ADDA adaptation restores performance to 0.91, closely matching source-domain baselines. Scale bar: 100 $\mu$m. “ns" and **** refer to p $\ge$ 0.05 and <0.0001, respectively.
• Figure 2: SIT-ADDA improves cross-modal reconstruction and preserves subcellular detail. (A–C) Ablation results indicate that keeping deeper, semantically meaningful blocks frozen and fine-tuning only the first three convolutional layers (SIT-ADDA) gives the best balance of stability and adaptability — outperforming conventional ADDA (all 16 layers trainable) across Pearson correlation (A), PSNR (B), and SSIM (C). (D) Representative examples of CD29 fluorescence prediction show that conventional ADDA can introduce cross-shaped and other hallucinated artifacts and fail to recover perinuclear enrichment, whereas SIT-ADDA more faithfully reconstructs perinuclear halo-like signal and preserves peripheral intensity. Scale bars: 100 $\mu$m. (E) Aggregate prediction accuracy across the test set: conventional ADDA raised Pearson correlation from a no-adaptation (N-a) baseline of 0.73 to 0.77, while SIT-ADDA further increased it to 0.85, reflecting superior recovery of fine-scale subcellular localization. (F) Two-dimensional UMAP embeddings of predicted images were computed for four conditions: source-trained (PC-PC and BF-BF), no adaptation (PC-BF), ADDA-adapted (ADDA), and SIT-ADDA-adapted (SIT-ADDA). SIT-ADDA predictions align more closely with the target-domain clusters (PC-PC and BF-BF), indicating improved correction of domain-induced distributional shifts. (G) Quantification via Euclidean distance between cluster centroids confirms this visual observation: SIT-ADDA predictions are consistently closer to ground-truth and same-modality outputs than those produced by conventional ADDA or by the no-adaptation baseline. *** and **** refer to p <0.001 and <0.0001, respectively.
• Figure 3: SIT-ADDA produces prediction improvements across common microscopy domain shifts. (A) Experimental design: we simulated three canonical domain perturbations, namely overexposure, spatial illumination gradient, and magnification (scaling), and evaluated adaptation performance by measuring reconstruction quality (Pearson correlation) across different perturbation strengths. (B–D) Overexposure results. (B) Summary curves show Pearson correlation as a function of increasing overexposure: SIT-ADDA consistently outperforms conventional ADDA and the no-adaptation baseline. (C) Representative image examples: extreme overexposure that produces saturation and clipping in baseline predictions is corrected by SIT-ADDA, restoring signal delineation and reducing hallucinated features. (D) Quantification: SIT-ADDA increased Pearson correlation by 7.8% compared to conventional ADDA, representing a statistically significant improvement. (E–G) Illumination-gradient results. (E) Summary curves for increasing gradient strength indicate greater recovery with SIT-ADDA. (F) SIT-ADDA normalizes spatial bias introduced by misaligned illumination, producing more uniform intensity profiles. (G) SIT-ADDA exceeded conventional ADDA by 13.2% in Pearson correlation. (H–J) Magnification (scaling) results. (H) Pearson versus scaling factor shows that both ADDA and SIT-ADDA achieve smaller gains compared to the other perturbation types. (I) SIT-ADDA can capture global morphology and some subcellular detail (e.g., perinuclear CD29 enrichment) in moderate scaling. (J) Although SIT-ADDA outperforms conventional ADDA by 14.3%, the source-domain U-Net already shows strong robustness to magnification shifts, yielding stable and reliable predictions. Scale bars: 100 $\mu$m. “ns" and **** refer to p $\ge$ 0.05 and <0.0001, respectively.
• Figure 4: Ensemble–based uncertainty enables unsupervised optimization of SIT-ADDA. (A) Workflow for estimating epistemic uncertainty. Five independently trained U-Net models were used to form an ensemble, with per-pixel variance across predictions serving as an approximation of predictive uncertainty. The mean yields the point estimate, while variance quantifies uncertainty without requiring ground truth. (B–E) Across all tested domain shifts, the scatter plots show a distinct separation between methods: SIT-ADDA (Layer 3, blue dots) predictions cluster tightly in the low-variance, high-accuracy region, whereas ADDA predictions (Layer 15, purple dots) spread into a regime of higher variance and lower accuracy. (F–I) Uncertainty inversely correlates with prediction accuracy. As standard deviation (Std) decreases, Pearson correlation systematically increases, confirming that predictive uncertainty is a reliable surrogate for adaptation performance.
• Figure 5: SIT-ADDA enables robust model transfer across diverse microscopy systems. (A) Experimental setup: models trained on Echo microscope data were tested on four distinct target platforms spanning a spectrum of optical performance, from lab-grade instruments to citizen-science tools, to evaluate the ability of SIT-ADDA to recover fluorescence-like predictions. (B) Mouse kidney tissue labeled with DAPI (nucleus, cyan) and phalloidin (actin, magenta) was used to assess adaptation performance. Scale bar: 500 $\mu$m. (C) Source domain reference: predictions from the Echo-trained model closely matched ground truth, accurately reconstructing actin filaments and nuclear morphology. (D) Feature-space analysis revealed five distinct and well-separated clusters, corresponding to the Echo source images and the four target platforms, validating pronounced distributional shifts. (E–H) Representative inputs and predictions for each target platform. Without adaptation, predictions failed to capture meaningful nuclear or cytoskeletal structure. In contrast, SIT-ADDA consistently restored biologically plausible features. Scale bars: 100 $\mu$m.