Interpretable deep learning illuminates multiple structures fluorescence imaging: a path toward trustworthy artificial intelligence in microscopy

Defu Yang; Ju Zhang; Luhong Jin; Mingyang Chen; Xuwei Xuan; Yun Cheng

REVIEW 5 major objections 6 minor 50 references

Interpretable deep learning illuminates multiple structures fluorescence imaging: a path toward trustworthy artificial intelligence in microscopy

T0 review · 5 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read AEMS-Net predicts two organelle images from one stain

desk verdict A useful architectural and interpretability contribution whose core replacement claim is undercut by synthetic-only quantitative validation. read the letter →

arxiv 2501.05490 v1 pith:324I6A54 submitted 2025-01-09 q-bio.SC cs.AIeess.IV

Mingyang Chen , Luhong Jin , Xuwei Xuan , Defu Yang , Yun Cheng , Ju Zhang This is my paper · ORCID

classification q-bio.SCcs.AIeess.IV

keywords interpretabledeeplearningfluorescencemicroscopysubcellularstructureseparationKolmogorov-ArnoldNetworksmitochondriamicrotubuleslive-cellimagingattentionmechanism

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a deep network, AEMS-Net, can take a single fluorescence image in which mitochondria and microtubules both appear in one channel and output separate, faithful images of the two structures, replacing the usual two-channel sequential acquisition with one channel plus computation. If true, live-cell studies of mitochondrial-microtubule interactions would need half the staining rounds and half the light exposure, lowering photobleaching and phototoxicity while allowing real-time tracking. The authors report that AEMS-Net beats a comparable U-Net by over 30% on imaging-quality metrics, and that Kolmogorov-Arnold layers plus Layer-CAM-style heatmaps show why the model separates structures the way it does. The central promise is better reconstruction together with inspectable, trustworthy AI for microscopy.

What carries the argument

The central mechanism is the KAN convolution layer, which splits the input into two branches: one applies a radial-basis-function transform followed by spline convolution, the other applies standard 2D convolution with SiLU activation, and the outputs are concatenated. This realizes the Kolmogorov-Arnold representation theorem in the network and gives each edge a learnable univariate function. The brightness adaptation layer after the final upsampling normalizes staining-intensity mismatch, the attention module prioritizes structure-relevant features, and residual connections preserve gradient flow. The interpretability machinery is a modified Layer-CAM that stores feature maps, computes gradients, and adds inverse-gradient analysis so both attended and ignored regions are visible.

What would settle it

Collect true co-labeled images by staining EMTB-3×eGFP microtubules and MitoTracker Orange mitochondria in the same cells, run AEMS-Net on the single-channel image, and compare its two predicted channels with the actual two-channel recordings; if PSNR and SSIM drop toward the U-Net baseline or mitochondrial signal bleeds into the microtubule channel, the central claim fails.

Watch

Extended reading notes

Core claim

AEMS-Net is a U-Net-style encoder-decoder in which standard convolution blocks are replaced by Kolmogorov-Arnold convolutions, so the network learns univariate functions along edges instead of fixed activations with learned weights. An attention mechanism and a brightness adaptation layer handle intensity differences between the two structures, and a modified Layer-CAM backpropagation produces positive and negative activation maps. The paper claims this combination separates mitochondria and microtubules with higher PSNR, lower NRMSE, and higher SSIM than a structurally identical U-Net, and that its decisions align with expert morphology: elongated features for microtubules, spherical and punctate features for mitochondria. On live-cell time series, AEMS-Net preserves thin microtubules and continuous mitochondria where U-Net loses or fragments them, and it tracks mitochondria moving along microtubules. The authors present this as a step toward trustworthy AI in microscopy, with the goal of cutting multi-staining side effects in half.

Load-bearing premise

The training data are synthetic overlays of mitochondrial and microtubule images taken from different cells, and the whole method rests on the assumption that these overlays faithfully represent a real single-color image in which both structures are labeled in the same cell.

Editorial extensions

If this is right

A single staining round and one acquisition channel would suffice for simultaneous mitochondrial and microtubule imaging, halving the light dose and staining burden for this structure pair.
Time-lapse records of interactions such as mitochondria moving along microtubules can be produced from individual frames without sequential channel switching.
The architecture can be retrained for further subcellular structures, so a five-structure study could in principle run with one staining step instead of five.
Modified Layer-CAM heatmaps give per-layer insight into what drives each separation, letting users see and correct failure modes before trusting the output.
The real-time inference speed fits event-driven microscopy, where acquisition is triggered by detected biological events.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct validation on true co-labeled images, where both structures are stained in the same cell, would test whether the synthetic overlay training pairs match reality; without it, the reported test-set gains may not transfer to ordinary lab acquisition.
The reported gain over U-Net may partly reflect the added brightness adaptation layer; an ablation that adds this layer to the U-Net baseline would separate the contribution of KAN from the contribution of brightness handling.
The interpretability is passive and post-hoc; building the same transparency into active, self-correcting models is a natural next step that the paper leaves open.
If the learned KAN functions are readable as intensity-to-output curves, biologists could extract quantitative descriptions of how fluorescence intensity drives each structure's reconstruction.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick read: this is a solid incremental engineering paper, not the paradigm shift the abstract promises. The new thing is AEMS-Net, a U-Net variant with KAN-style convolutions, attention, and a brightness-adaptive layer, plus a modified Layer-CAM interpretability workflow. On synthetic test overlays of mitochondria and microtubule images from different cells, it beats a vanilla U-Net by a wide margin, and the attention/BAL ablations confirm both modules earn their keep. That's real work, and the interpretability analysis is a useful addition to a literature that mostly ignores failure modes.

The soft spot is exactly the one the stress-test note flags. Training, validation, and test pairs are all built by superimposing images of independently stained cells. The real co-labeled live-cell data is only evaluated qualitatively (Fig. 3-4, videos), so the headline claim — that one acquisition plus computation replaces sequential double-color imaging — is not quantitatively established for real simultaneous acquisition, where crosstalk and same-cell spatial correlations change the conditional distribution. The paper is not shy about the synthetic construction, but the abstract and discussion overreach beyond it.

A few smaller issues: no comparison to the closest prior method (Jin et al. 2024, ref 13, which also separates two structures from one label), no significance tests on the metric deltas, and the KAN component is never ablated separately — the baseline differs in three ways at once. The "30%" claim is relative PSNR, which is logarithmic, so it oversells the perceptual gain. And the GitHub link in the text is empty, which conflicts with the data/code availability statement.

Who this is for: people working on deep learning for fluorescence microscopy who want an interpretability recipe and a somewhat better architecture for structure separation. It deserves a serious referee, but revision should demand either a quantitative real co-labeled test set (even a small one) or a scaled-back claim, plus a comparison to Jin et al. and release of code.

Referee Report

5 major / 6 minor

Summary. The paper introduces AEMS-Net, a U-Net variant augmented with Kolmogorov-Arnold convolutions, attention modules, and a brightness adaptation layer, intended to reconstruct mitochondrial and microtubule images from a single fluorescence image. Training and test data are generated by superimposing independently stained mitochondrial and microtubule images from different cells, with the two component images serving as ground truth. Quantitative evaluation on this synthetic test set reports PSNR, NRMSE, and SSIM improvements over a vanilla U-Net, and qualitative results are shown on a live-cell application dataset of co-transfected cells. The authors claim over 30% imaging-quality improvement and interpretability through a modified Layer-CAM analysis.

Significance. If the central claim were fully validated, the method could reduce acquisition time, photobleaching, and phototoxicity in live-cell multi-structure imaging by replacing sequential multi-color acquisition with a single-color acquisition plus computation. The paper provides source data files and promises public code, which facilitates reproducibility. The interpretability analysis using Layer-CAM is a useful contribution to reliability assessment in microscopy deep learning. However, the significance is currently limited by the gap between the synthetic training/test regime and real co-labeled imaging, and by the absence of quantitative comparison with the closest prior method.

major comments (5)

[Data Preprocessing; Results; Figs. 3-4; Supplementary Tables 3-5] The central claim of replacing sequential double-color imaging with single-color acquisition is not quantitatively validated on real co-labeled images. All training and test pairs are synthetic overlays of independently stained mitochondrial and microtubule images from different cells, and all reported PSNR/NRMSE/SSIM improvements are computed on this synthetic test distribution. The live-cell Application dataset is real co-transfected data, but evaluation there is purely qualitative (Figs. 3-4), with no comparison against measured ground-truth channels. The authors should provide quantitative validation on real single-color images of co-labeled cells, for example by acquiring sequential dual-color ground truth on the same cells, or by demonstrating that the synthetic overlay statistics match real co-labeled images.
[Table 2; Ablation Studies] The ablation study only toggles the Attention module and the Brightness Adaptation Layer while keeping KAN convolutions fixed in all conditions. The KAN-specific contribution is therefore never isolated: a baseline with standard convolutions plus Attention and BAL is missing. Since the title and introduction attribute the method's performance and interpretability partly to the Kolmogorov-Arnold representation theorem, the authors must ablate KAN to show that it adds value beyond the attention and brightness components.
[Introduction; Results; Supplementary Tables 3-5] The only deep-learning comparator is a vanilla U-Net. The closest prior method, Jin et al. (ref 13), which also separates multiple structures labeled with the same fluorophore, is discussed in the Introduction but never quantitatively compared. The headline claim of 'over 30% improvement compared to traditional deep learning methods' is therefore not positioned against the state of the art for the same task. A quantitative comparison with Jin et al. or a detailed justification of why that method is not applicable is needed.
[Loss Function and Training Details, Eqs. (1)-(6)] The loss function weights w_mse, w_grad, w_con, w_focal and the hyperparameters alpha, gamma, and T are introduced in Eqs. (2)-(6) but no numerical values are reported anywhere in the paper or supplementary material, and no sensitivity analysis is provided. Reproduction is impossible without these values. Additionally, Eq. (4) defines the gradient loss as a signed sum of gradient differences without squaring or absolute value, which can cancel across pixels; this appears to be a typo and should be corrected to a proper gradient-matching term.
[Results: Enhanced Performance; Supplementary Tables 3-5; Fig. 2g] Reported metric improvements are not accompanied by any statistical significance tests. For the main test-set comparisons the mean differences are large, but for the ablation study in Table 2, the Attention-only and Attention+BAL conditions differ by less than 1 dB in PSNR (e.g., 27.1±2.7 vs 28.03±3.5 for mitochondria), and it is unclear whether this difference is significant. The authors should provide paired significance tests or confidence intervals across test images, especially for the ablation comparisons that support the claim of synergistic gains.

minor comments (6)

[Abstract and Introduction] The abstract says the method requires 'a single staining procedure,' but the live-cell application uses co-transfection with two fluorescent proteins (EMTB-3×eGFP and Tom20-mEmerald). The method reduces the number of spectral channels, not the number of stains; please rephrase to 'single-color acquisition' or similar.
[Eq. (5)] The label for the contrastive loss in Eq. (5) is garbled as 'L_mtntmtmtcrcr' and should be 'L_con'; the equation also has an unmatched parenthesis in the denominator.
[Evaluation Metrics, Eqs. (7)-(9)] The NRMSE definition in Eq. (8) differs from the standard normalized root mean square error (which normalizes by the range of values); here the denominator is the square root of the sum of squares of the ground truth. This should be clarified, and consistency with the skimage implementation should be checked.
[Data Preprocessing] The 'comprehensive intra-group image superimposition analyses' procedure is not specified in enough detail: the weighting of the two channels, normalization before summation, random pairing scheme, and any augmentation steps are not described. Please provide the exact overlay algorithm for reproducibility.
[Supplementary Figures] Supplementary Figure numbering is inconsistent: both Supplementary Figure 1 and Supplementary Figure 3 are titled 'AEMS-Net architecture details,' and the captions for Supplementary Figures 2 and 3 appear to be swapped. Please renumber and correct the captions.
[Data and Code Availability] The GitHub link is missing from the text ('available at the following GitHub link:' with no URL). Please provide the actual repository address.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: AEMS-Net is trained and evaluated on held-out synthetic overlays, and the only author self-citations are background, not load-bearing.

full rationale

The paper's central quantitative claim is a supervised image-to-image reconstruction benchmark: training pairs are synthetic overlays of independently stained mitochondria and microtubule images, with the pre-overlay components as ground truth (Data Preprocessing). The reported PSNR/NRMSE/SSIM improvements are computed on a held-out test set of the same synthetic-overlay type, under identical data splits for AEMS-Net and U-Net. No metric is computed by fitting a parameter to the test set and then renaming that fit as a prediction; the comparison is a standard held-out evaluation. The live-cell Application dataset is real co-transfected imaging, but evaluation there is qualitative and does not enter the numerical improvement claim, so it is an external-validity limitation rather than a circular step. The only overlapping self-citations are Jin et al. (refs 12 and 13), where co-author Luhong Jin appears; these are cited as prior work and contrast, not as the justification for AEMS-Net's architecture or as a uniqueness proof. The Kolmogorov-Arnold representation theorem is an external mathematical result, and its use through KAN convolution is an architectural choice, not a result derived from the paper's own outputs. No equation in the paper reduces by construction to its own input, and no fitted value is renamed as a prediction. The main caveat—that synthetic overlays of different cells may not faithfully represent true co-labeled same-cell images—is a question of dataset realism and generalization, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four free-parameter groups (loss weights, focal loss, contrastive temperature, KAN grid settings) and on four assumptions, the most fragile being that synthetic overlays of images from different cells are equivalent to real co-labeled single-color images. No new physical entities are introduced.

free parameters (4)

Loss component weights w_mse, w_grad, w_con, w_focal = Not stated
Equation (2) defines organelle loss as a weighted sum; no values or tuning protocol are reported, and they directly affect reconstruction quality.
Focal loss hyperparameters alpha and gamma = Not stated
Equation (6) includes alpha and gamma; standard choices are common but no values are given, affecting separation of dim structures.
Contrastive loss temperature T = Not stated
Equation (5) includes temperature T; no value reported.
KAN grid and spline parameters = Not stated
ImprovedFastKAN is described qualitatively (RBF, spline convolution, grid-based centroid modeling) but grid size, spline order, and basis count are omitted; these affect capacity.

assumptions (4)

standard math Kolmogorov-Arnold representation theorem
Invoked in Abstract and Methods to motivate KAN convolutions. The theorem is true but is used as architectural inspiration, not as a formal guarantee of the network's representational or explainability properties.
domain assumption Synthetic overlays of independently stained organelles from different cells are equivalent to single-color co-labeled images of both organelles in one cell
Data Preprocessing builds training pairs by superimposing separately acquired mitochondrial and microtubule images; the entire supervised claim depends on this equivalence holding at inference on real single-color data.
domain assumption Fluorescence intensities from two structures combine additively in the overlay
The overlay operation is not precisely defined (sum, max, or weighted combination), yet the network is trained to invert it; the inversion task changes if the true combination is nonlinear or includes crosstalk.
domain assumption COS-7 cell morphologies and imaging conditions generalize to other live-cell applications
All training, validation, and test images come from COS-7 cells on an Olympus FV3000; the discussion claims broader applicability without cross-cell-type validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Interpretable deep learning illuminates multiple structures fluorescence imaging: a path toward trustworthy artificial intelligence in microscopy." pith.science (2026). https://pith.science/paper/324I6A54

@misc{pith2026250105490,
  author       = {Pith},
  title        = {Pith review of: Interpretable deep learning illuminates multiple structures fluorescence imaging: a path toward trustworthy artificial intelligence in microscopy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/324I6A54}},
  note         = {Machine review of arXiv:2501.05490}
}

read the original abstract

Live-cell imaging of multiple subcellular structures is essential for understanding subcellular dynamics. However, the conventional multi-color sequential fluorescence microscopy suffers from significant imaging delays and limited number of subcellular structure separate labeling, resulting in substantial limitations for real-time live-cell research applications. Here, we present the Adaptive Explainable Multi-Structure Network (AEMS-Net), a deep-learning framework that enables simultaneous prediction of two subcellular structures from a single image. The model normalizes staining intensity and prioritizes critical image features by integrating attention mechanisms and brightness adaptation layers. Leveraging the Kolmogorov-Arnold representation theorem, our model decomposes learned features into interpretable univariate functions, enhancing the explainability of complex subcellular morphologies. We demonstrate that AEMS-Net allows real-time recording of interactions between mitochondria and microtubules, requiring only half the conventional sequential-channel imaging procedures. Notably, this approach achieves over 30% improvement in imaging quality compared to traditional deep learning methods, establishing a new paradigm for long-term, interpretable live-cell imaging that advances the ability to explore subcellular dynamics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 46 canonical work pages

[1]

Lee, R. G. et al. Quantitative subcellular reconstruction reveals a li pid mediated inter -organelle biogenesis network. Nat. Cell Biol. 26, 57– 71 (2024)

work page 2024
[2]

Tsukui, T., Wolters, P. J. & Sheppard, D. Alveolar fibroblast lineag e orchestrates lung inflammation and fibrosis. Nature 631, 627–634 (2024)

work page 2024
[3]

Mills, C. E. et al. Multiplexed and reproducible high content scree ning of live and fixed cells using Dye Drop. Nature Communications 13, 6 918 (2022)

work page 2022
[4]

Garini, Y ., Young, I. T. & McNamara, G. Spectral imaging: Princip les and applications. Cytometry Part A 69A, 735–747 (2006)

work page 2006
[5]

Giepmans, B. N. G., Adams, S. R., Ellisman, M. H. & Tsien, R. Y . The Fluorescent Toolbox for Assessing Protein Location and Function. S cience 312, 217–224 (2006)

work page 2006
[6]

Shaner, N. C. et al. Improving the photostability of bright monome ric orange and red fluorescent proteins. Nature Methods 5, 545–551 (2008)

work page 2008
[7]

Waters, J. C. Accuracy and precision in quantitative fluorescence m icroscopy. Journal of Cell Biology 185, 1135–1148 (2009)

work page 2009
[8]

Carlton, P. M. et al. Fast live simultaneous multiwavelength four -di mensional optical microscopy. Proceedings of the National Academy of Scie nces 107, 16016–16022 (2010)

work page 2010

Show all 50 references

[9]

Icha, J., Weber, M., Waters, J. C. & Norden, C. Phototoxicity in li ve fluorescence microscopy, and how to avoid it. BioEssays 39, 1700003 (2017)

2017
[10]

& Sirmacek, B

Vinuesa, R. & Sirmacek, B. Interpretable deep- learning models to help achieve the sustainable development goals. Nat. Mach. Intell. 3, 926–9 26 (2021)

2021
[11]

Achtibat, R. et al. From attribution maps to human- understandable explanations through Concept Relevance Propagation. Nature Machine Intelli gence 5, 1006–1019 (2023)

2023
[12]

Jin, L. et al. Deep learning enables structured illumination microsc opy with low light levels and enhanced speed. Nat. Commun. 11, 1934 (20 20)

1934
[13]

Jin, L. et al. Deep learning permits imaging of multiple structures with the same fluorophores. Biophys. J. 123, 3540–3549 (2024)

2024
[14]

Liao, J. et al. Deep-MSIM: Fast Image Reconstruction with Deep Learning in Multifocal Structured Illumination Microscopy. Advanced Scienc e 10, 2300947 (2023)

2023
[15]

& Brox, T

Ronneberger, O., Fischer, P. & Brox, T. U -net: convolutional netwo rks for biomedical image segmentation. in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (eds. Navab, N., Hornegger, J., Wells, W. M. & Frangi, A. F.) 234–241 (Springer International ...

2015 doi
[16]

Bell, R. D. et al. Automated multi- scale computational pathotyping (AMSCP) of inflamed synovial tissue. Nature Communications 15, 7503 (20 24)
[17]

Li, Z. et al. D-LMBmap: a fully automated deep -learning pipeline for whole -brain profiling of neural circuitry. Nature Methods 20, 1593–1604 (2023)

2023
[18]

Zargari, A. et al. DeepSea is an efficient deep -learning model for single-cell segmentation and tracking in time -lapse microscopy. Cell Reports Methods 3, (2023)

2023
[19]

Chen, R. et al. Single-frame deep -learning super -resolution microsc opy for intracellular dynamics imaging. Nature Communications 14, 2854 (2 023)
[20]

Chaudhary, S., Moon, S. & Lu, H. Fast, efficient, and accurate ne uro-imaging denoising via supervised deep learning. Nature Communications 13, 5165 (2022)

2022
[21]

Zhang, Y . et al. Rapid detection of neurons in widefield calcium i maging datasets after training with synthetic data. Nature Methods 20, 747– 754 (2023)

2023
[22]

Qu, L. et al. Self-inspired learning for denoising live -cell super -res olution microscopy. Nature Methods 21, 1895– 1908 (2024)

2024
[23]

Saguy, A. et al. DBlink: dynamic localization microscopy in super spatiotemporal resolution via deep learning. Nature Methods 20, 1939–1948 (2023)

2023
[24]

Chen, Y .-I. et al. Generative adversarial network enables rapid and robust fluorescence lifetime image analysis in live cells. Communications Bi ology 5, 18 (2022)

2022
[25]

Ma, J. et al. The multimodality cell segmentation challenge: toward universal solutions. Nature Methods 21, 1103– 1113 (2024)

2024
[26]

& Barbastathi s, G

Pham, T., Boquet -Pujadas, A., Mondal, S., Unser, M. & Barbastathi s, G. Deep -prior ODEs augment fluorescence imaging with chemical sensor s. Nature Communications 15, 9172 (2024)

2024
[27]

& Torralba, A

Zhou, B., Khosla, A., Lapedriza, A., Oliva, A. & Torralba, A. Lea rning deep features for discriminative localization. in 2921–2929 (2016)

2016
[28]

Selvaraju, R. R. et al. Grad-CAM: visual explanations from deep n etworks via gradient -based localization. in 618 –626 (2017)

2017
[29]

& Balasubramanian, V

Chattopadhay, A., Sarkar, A., Howlader, P. & Balasubramanian, V . N. Grad -CAM++: generalized gradient -based visual explanations for deep co nvolutional networks. in 2018 IEEE Winter Conference on Applications of Computer Vision (WACV) 839–847 (2018). doi:10.1109/WACV .2018.00097

2018
[30]

-B., Hou, Q., Cheng, M.- M

Jiang, P.- T., Zhang, C. -B., Hou, Q., Cheng, M.- M. & Wei, Y . Laye rCAM: exploring hierarchical class activation maps for localization. IEEE T rans. Image Process. 30, 5875–5888 (2021)

2021
[31]

Xu, R.- Z. et al. Attention improvement for data -driven analyzing fl uorescence excitation -emission matrix spectra via interpretable attention mec hanism. npj Clean Water 7, 73 (2024)

2024
[32]

Li, H. et al. CGMega: explainable graph neural network framework with attention mechanisms for cancer gene module dissection. Nature Com munications 15, 5997 (2024)

2024
[33]

Angello, N. H. et al. Closed-loop transfer enables artificial intellige nce to yield chemical knowledge. Nature 633, 351–358 (2024)

2024
[34]

Cortacero, K. et al. Evolutionary design of explainable algorithms f or biomedical image segmentation. Nature Communications 14, 7112 (2023)

2023
[35]

Seo, B. & Li, J. Explainable machine learning by SEE -Net: closin g the gap between interpretable models and DNNs. Scientific Reports 14, 2 6302 (2024)

2024
[36]

Miranda, F. et al. Interpretable artificial intelligence for classificatio n of alveolar bone defect in patients with cleft lip and palate. Scientific Re ports 13, 15861 (2023)

2023
[37]

Lee, M. et al. Quantifying arousal and awareness in altered states of consciousness using interpretable deep learning. Nature Communications 13, 1064 (2022)

2022
[38]

& Tiwary, P

Mehdi, S. & Tiwary, P. Thermodynamics -inspired explanations of a rtificial intelligence. Nature Communications 15, 7859 (2024)

2024
[39]

Liu, Z. et al. KAN: Kolmogorov -Arnold Networks. Preprint at http s://doi.org/10.48550/arXiv.2404.19756 (2024)

2024 doi arXiv
[40]

Kolmogorov- Arnold Networks are Radial Basis Function Net works

Li, Z. Kolmogorov- Arnold Networks are Radial Basis Function Net works. Preprint at https://doi.org/10.48550/arXiv.2405.06721 (2024)

2024 doi arXiv
[41]

& Sun, J

He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for i mage recognition. in 770–778 (2016)

2016
[42]

Barredo Arrieta, A. et al. Explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI. I nf. Fusion 58, 82–115 (2020)

2020
[43]

Liang, J. et al. Deep learning supported discovery of biomarkers fo r clinical prognosis of liver cancer. Nat. Mach. Intell. 5, 408–420 (2023)

2023
[44]

Zhou, L. et al. An interpretable deep learning workflow for discov ering subvisual abnormalities in CT scans of COVID -19 inpatients and surv ivors. Nat. Mach. Intell. 4, 494–503 (2022)

2022
[45]

Luo, A. et al. Visual interpretable MRI fine grading of meniscus i njury for intelligent assisted diagnosis and treatment. npj Digital Med. 7, 9 7 (2024)

2024
[46]

Barnett, A. J. et al. A case- based interpretable deep learning model for classification of mass lesions in digital mammography. Nat. Mach. Inte ll. 3, 1061–1070 (2021)

2021
[47]

Durmaz, A. R. et al. A deep learning approach for complex micro structure inference. Nature Communications 12, 6272 (2021)

2021
[48]

Mahecic, D. et al. Event-driven acquisition for content -enriched mi croscopy. Nat. Methods 19, 1262–1267 (2022)

2022
[49]

& Zhuang, X

Gao, S., Zhou, H., Gao, Y . & Zhuang, X. BayeSeg: bayesian mod eling for medical image segmentation with interpretable generalizability. Me d. Image Anal. 89, 102889 (2023)

2023
[50]

Wang, Y . et al. Kolmogorov Arnold Informed neural network: A p hysics-informed deep learning framework for solving forward and inverse pr oblems based on Kolmogorov Arnold Networks. Computer Methods in Appl ied Mechanics and Engineering 433, 117518 (2025). Supplementary Table...

2025

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Lee, R. G. et al. Quantitative subcellular reconstruction reveals a li pid mediated inter -organelle biogenesis network. Nat. Cell Biol. 26, 57– 71 (2024)

work page 2024

[2] [2]

Tsukui, T., Wolters, P. J. & Sheppard, D. Alveolar fibroblast lineag e orchestrates lung inflammation and fibrosis. Nature 631, 627–634 (2024)

work page 2024

[3] [3]

Mills, C. E. et al. Multiplexed and reproducible high content scree ning of live and fixed cells using Dye Drop. Nature Communications 13, 6 918 (2022)

work page 2022

[4] [4]

Garini, Y ., Young, I. T. & McNamara, G. Spectral imaging: Princip les and applications. Cytometry Part A 69A, 735–747 (2006)

work page 2006

[5] [5]

Giepmans, B. N. G., Adams, S. R., Ellisman, M. H. & Tsien, R. Y . The Fluorescent Toolbox for Assessing Protein Location and Function. S cience 312, 217–224 (2006)

work page 2006

[6] [6]

Shaner, N. C. et al. Improving the photostability of bright monome ric orange and red fluorescent proteins. Nature Methods 5, 545–551 (2008)

work page 2008

[7] [7]

Waters, J. C. Accuracy and precision in quantitative fluorescence m icroscopy. Journal of Cell Biology 185, 1135–1148 (2009)

work page 2009

[8] [8]

Carlton, P. M. et al. Fast live simultaneous multiwavelength four -di mensional optical microscopy. Proceedings of the National Academy of Scie nces 107, 16016–16022 (2010)

work page 2010