{"id":"363315da-d7eb-445f-a757-8ccee13de482","arxiv_id":"2501.05490","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AEMS-Net, a U-Net variant with KAN convolutions, attention, and brightness adaptation, reconstructs mitochondrial and microtubule images from one fluorescence image and reports large gains over vanilla U-Net.","lead":"A deep learning model named AEMS-Net separates two subcellular structures (mitochondria and microtubules) from a single fluorescence image, reducing the need for multiple staining rounds. Biologists could image interactions between organelles in real time with fewer side effects from photobleaching and phototoxicity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central replacement claim rests on synthetic overlays of independently stained cells standing in for real co-labeled images; reported gains are measured on synthetic test data, so transfer to real simultaneous acquisition is unvalidated.","rationale":"The reader's weakest assumption correctly identifies the synthetic-overlay validity as the load-bearing premise for the central claim. I agree with that identification, and the paper's own text supports it: training inputs are made by superimposing independently stained mitochondrial and microtubule images from different cells, and the reported quantitative gains are all on that synthetic test set. The Application dataset is real co-labeled imaging, but it is evaluated only qualitatively, so it does not currently close the gap. I also note positive aspects: the ablation study is informative, the leakage-prevention separation by cell is a good practice, and the qualitative live-cell videos provide some evidence of practical utility. However, the core quantitative claim of replacing double-color acquisition has not been validated on real co-labeled images with known ground truth. This does not overturn the manuscript; it strengthens the need for the conditional verdict the reader already assigned. A single targeted experiment with real co-labeled pairs would settle whether the synthetic training distribution is faithful enough.","tokens_in":15547,"tokens_out":3896,"duration_ms":42592,"concrete_test":"Acquire a held-out set of at least 20 COS-7 cells co-labeled for microtubules (EMTB-3×eGFP) and mitochondria, acquiring both channels simultaneously with the same FV3000 setup. Feed only the microtubule channel into the trained AEMS-Net and compare its predicted mitochondrial channel against the measured mitochondrial channel using PSNR, NRMSE, and SSIM; compute the same metrics on the synthetic overlay test set. If the real-image metrics are substantially worse than the synthetic metrics (e.g., a relative PSNR drop greater than 20%), the synthetic-overlay training distribution does not transfer to real co-labeled acquisition, and the central replacement claim fails. If the metrics are comparable, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims that AEMS-Net can replace double-color sequential imaging with single-color acquisition plus computation, with over 30% imaging-quality improvement. The training and test pipelines, however, are built by 'comprehensive intra-group image superimposition analyses' (Data Preprocessing): inputs are overlays of mitochondrial and microtubule images taken from different, independently stained cells, and the ground-truth outputs are the pre-overlay component images. This makes the learning task one of inverting a synthetic sum of independent fields. In a real co-labeled cell, the two structures are not independent: mitochondria move along microtubules, so the same pixel can contain both signals, and the two channels share acquisition crosstalk and same-cell spatial correlations. The network may therefore be learning to invert an artificial transformation whose conditional distribution differs from the true acquisition process. The reported PSNR/NRMSE/SSIM improvements (Supplementary Tables 3-5) are computed on this synthetic test regime. The live-cell Application dataset is real co-labeled imaging, but evaluation there is qualitative (Figs. 3-4, Supplementary Videos), with no quantitative comparison against measured ground-truth channels. If the synthetic overlay distribution does not match real co-labeled images, the headline claim of replacing sequential double-color imaging with single-color acquisition is not established, even though the architecture may perform well on synthetic superimposition data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":15851,"tokens_out":5488,"duration_ms":55408,"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":[{"comment":"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.","section":"Data Preprocessing; Results; Figs. 3-4; Supplementary Tables 3-5"},{"comment":"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.","section":"Table 2; Ablation Studies"},{"comment":"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.","section":"Introduction; Results; Supplementary Tables 3-5"},{"comment":"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.","section":"Loss Function and Training Details, Eqs. (1)-(6)"},{"comment":"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.","section":"Results: Enhanced Performance; Supplementary Tables 3-5; Fig. 2g"}],"minor_comments":[{"comment":"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.","section":"Abstract and Introduction"},{"comment":"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.","section":"Eq. (5)"},{"comment":"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.","section":"Evaluation Metrics, Eqs. (7)-(9)"},{"comment":"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.","section":"Data Preprocessing"},{"comment":"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.","section":"Supplementary Figures"},{"comment":"The GitHub link is missing from the text ('available at the following GitHub link:' with no URL). Please provide the actual repository address.","section":"Data and Code Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practical problem in live-cell fluorescence microscopy, and the architectural innovations are reasonable. However, the central replacement claim is currently supported only by synthetic test-set metrics, with the real-data validation being qualitative. The missing KAN ablation and the absence of a comparison with the closest prior method further weaken the quantitative claims. These issues are fixable with additional experiments and analysis, so I recommend major revision rather than rejection. The manuscript would benefit from a clearer articulation of what exactly is 'single staining' versus 'single-color acquisition,' and from a more careful statistical treatment of the metric comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nA 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.\n\nWho 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.","headline":"A useful architectural and interpretability contribution whose core replacement claim is undercut by synthetic-only quantitative validation.","tokens_in":16361,"tokens_out":2467,"would_cite":false,"duration_ms":24300,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AEMS-Net predicts two organelle images from one stain","keywords":["interpretable deep learning","fluorescence microscopy","subcellular structure separation","Kolmogorov-Arnold Networks","mitochondria","microtubules","live-cell imaging","attention mechanism"],"falsifier":"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.","tokens_in":15364,"feed_emoji":"🔬","tokens_out":9178,"duration_ms":81748,"temperature":0.7,"pith_summary":"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.","feed_headline":"Network splits one stained image into two organelle channels","feed_subtitle":"One fluorescence channel plus AEMS-Net replaces two-channel sequential acquisition, halving staining and light exposure for live cells.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the U-Net architecture that serves as the baseline and the backbone that AEMS-Net modifies.","marker":"[15]"},{"why":"Supplies the Kolmogorov-Arnold representation theorem and the KAN design that AEMS-Net implements in its convolution layers.","marker":"[39]"},{"why":"Gives the radial-basis-function formulation of KAN that underlies the spline-convolution branch.","marker":"[40]"},{"why":"Earlier work on separating multiple structures labeled with the same fluorophore that AEMS-Net extends to single-stain multi-structure prediction.","marker":"[13]"},{"why":"Layer-CAM is the class-activation method the authors modify for per-layer interpretability of AEMS-Net.","marker":"[30]"},{"why":"Event-driven acquisition is the intelligent-microscopy workflow the authors say real-time AEMS-Net inference can feed.","marker":"[48]"}],"fun_headline_variants":["One stained image yields two organelle maps","AEMS-Net splits one channel into mitochondria and microtubules","Deep learning halves imaging steps for live-cell organelles","Interpretable network produces two structures from one frame","Double organelle insight from a single fluorescence capture"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["One stained image yields two organelle maps","AEMS-Net splits one channel into mitochondria and microtubules","Deep learning halves imaging steps for live-cell organelles","Interpretable network produces two structures from one frame","Double organelle insight from a single fluorescence capture"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000841,"raw_usage":{"total_tokens":3656,"prompt_tokens":931,"completion_tokens":2725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":2652}},"tokens_in":547,"tokens_out":2725,"duration_ms":19709,"temperature":1.0,"reasoning_tokens":2652,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:19:58.320521+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier work on separating multiple structures labeled with the same fluorophore that AEMS-Net extends to single-stain multi-structure prediction."},{"cited_title":"-B., Hou, Q., Cheng, M.- M","cited_arxiv_id":null,"evidence_quote":"Layer-CAM is the class-activation method the authors modify for per-layer interpretability of AEMS-Net."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Event-driven acquisition is the intelligent-microscopy workflow the authors say real-time AEMS-Net inference can feed."}],"review_version":1}