{"id":"e8548422-d9d8-4eb6-a85e-5da17f77c461","arxiv_id":"2608.10807","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Implicit neural representations with per-eye latent vectors predict future and intermediate geographic atrophy lesion shape and size from longitudinal FAF images better than comparison models on a small cohort.","lead":"This paper trains a neural network to learn how geographic atrophy, a form of age-related blindness, grows in each patient's eye from a few scans taken months apart. It then generates missing scans and predicts future lesion size and shape, which could help doctors discuss prognosis.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Statistical basis for 'better than baselines' is fragile: n=6 test eyes, overlapping error bars, and a duplicated baseline row in Table 1.","rationale":"The reader's weakest_assumption targets the smooth-trajectory manifold hypothesis, which is a reasonable modeling limitation but is not directly falsified by the presented experiments. My stress-test instead focuses on the statistical foundation of the quantitative superiority claim, which is more immediately load-bearing: the evaluation uses only 6 test eyes, the error bars are large and overlapping, and no significance testing is reported. The duplicated Linear extrap. row in Table 1 is a concrete red flag that the numerical comparison may contain errors, as is the abstract's claim about FAF quality that the paper's own conclusion contradicts. These issues do not invalidate the conceptual contribution, and the qualitative trajectory plots are encouraging, but they must be addressed before the central claim can be taken at face value. The reader's CONDITIONAL verdict already captures this level of concern; therefore I leave the verdict unchanged.","tokens_in":8537,"tokens_out":6943,"duration_ms":74722,"concrete_test":"Compute paired differences (Ours minus Copy-forward) on the 6 test eyes for DICE and MAE in Scenario 2 and run a Wilcoxon signed-rank test or a paired bootstrap with 95% confidence intervals for the mean/median difference. Also independently recompute the Linear extrap. row from the submission's code or raw results; if it truly equals Copy-forward, rerun the baseline evaluation and verify the reported numbers. Report the resulting p-values and confidence intervals in a revised Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Scenario 2's headline numbers (DICE 0.91, MAE 0.20 mm2) come from only 6 test eyes. The reported standard deviations are large: MAE is 0.20±0.29 for Ours versus 0.38±0.24 for Copy-forward; DICE is 0.91±0.05 versus 0.86±0.09. No significance tests or confidence intervals are provided. A paired test on the 6 per-eye differences would likely show non-significance, meaning the central claim of superior progression prediction is not statistically established. This concern is compounded by Table 1 Scenario 2, where Linear extrap. has exactly the same metrics as Copy-forward (17.86/0.61/0.14/0.86/6.98/0.38), an implausible copy-paste error that casts doubt on the reliability of the numerical comparison. Additionally, the abstract claims 'without sacrificing FAF image quality', but Copy-forward outperforms Ours on PSNR, SSIM, and LPIPS in both Scenario 1 and 2, and the conclusion explicitly acknowledges that simple copying yields the best reconstruction. These are concrete, addressable problems; until resolved, the evidence for the headline claim is not yet convincing, even though the method is a plausible and interesting extension of INRs to longitudinal medical imaging.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a time-conditioned implicit neural representation (INR) with per-eye latent codes to jointly model Fundus Autofluorescence (FAF) image appearance and Geographic Atrophy (GA) segmentation over time. The method uses an auto-decoder MLP with SIREN activations, FiLM modulation conditioned on the latent vector, time in weeks, and patient age, and is split into a reconstruction head and a segmentation head. Test-time adaptation optimizes a new latent code on a subset of visits of an unseen eye while keeping the INR frozen, then predicts held-out past, future, or intermediate visits. The method is evaluated on the OMEGA longitudinal FAF dataset in three scenarios: single-pair extrapolation, test-time adaptation with full patient history, and missing-visit interpolation. The authors report competitive or best segmentation quality (DICE, Hausdorff distance, lesion-area MAE) and competitive FAF reconstruction, with code released publicly.","tokens_in":8878,"tokens_out":4220,"duration_ms":43857,"significance":"If the results hold, this is a useful extension of implicit neural representations to longitudinal ophthalmic imaging, enabling continuous, patient-specific disease trajectories from very limited data and providing both image-level and segmentation-level predictions. The paper has several strengths: it uses a patient-wise data split, evaluates on held-out visits without leaking the target visit into the latent optimization, provides an ablation of latent-grid configurations, and releases code. However, the evidence supporting the headline superiority claim is weakened by the very small test set (n=6), the absence of significance testing, and a clear numerical inconsistency in Table 1. The abstract overstates FAF image-quality performance relative to the paper's own quantitative results and conclusion. These issues are addressable, and the underlying modeling idea is plausible, so the contribution is worth considering after substantial revision.","major_comments":[{"comment":"The abstract states the method achieves the lowest MAE and highest DICE 'without sacrificing FAF image quality,' but Table 1 contradicts this: in Scenario 1, Ours has lower PSNR than Copy-forward (16.19 vs 16.49) and higher LPIPS (0.28 vs 0.16); in Scenario 2, Copy-forward achieves higher PSNR (17.86 vs 16.31), higher SSIM (0.61 vs 0.61, tied), and lower LPIPS (0.14 vs 0.26). The conclusion also explicitly acknowledges that copying the last image yields the best reconstruction quality. The abstract should be revised to report the segmentation advantages only, or to qualify the FAF claim consistently with the presented numbers.","section":"Abstract and Table 1"},{"comment":"The row for 'Linear extrap. [12]' reports exactly the same values as 'Copy-forward' (PSNR 17.86, SSIM 0.61, LPIPS 0.14, DICE 0.86, HD 6.98, MAE 0.38). This is implausible as a real result and appears to be a copy-paste error. Because these rows are used to support the claim that the method outperforms classical baselines, the numerical integrity of the comparison must be verified and corrected before the results can be assessed.","section":"Table 1, Scenario 2"},{"comment":"The central claim that the method is better than the baselines rests on only six test eyes, and no significance tests or confidence intervals are provided. For Scenario 2, the reported DICE is 0.91±0.05 versus 0.86±0.09 for Copy-forward, and the lesion-area MAE is 0.20±0.29 versus 0.38±0.24; these standard deviations overlap considerably at n=6. The authors should report paired tests (e.g., Wilcoxon signed-rank or paired t-test) with effect sizes, or otherwise provide per-eye differences, for DICE, MAE, and HD in all scenarios. Without this, the assertion of superior progression prediction is not statistically established.","section":"Section 4 and Table 1 (statistical evaluation)"},{"comment":"T-I2SBUNet [16] is listed as a comparison method and appears in Scenario 1, but it is absent from Scenario 2 and from the Missing-visits table. If the comparison is intended to cover all scenarios, the omission should be explained, or the method should be included. This is relevant because the qualitative comparison in Fig. 2 may not represent all methods in all settings.","section":"Table 1, Scenario 2 and Missing visits"}],"minor_comments":[{"comment":"The sentence 'The hyperparameters were optimised on the validation set' would benefit from a full list of searched hyperparameter ranges and the selection criterion; otherwise the reproducibility of the configuration choice is limited.","section":"Section 3, Implementation Details"},{"comment":"The caption states the red box is the best configuration 'by subjective judgment over all five metrics.' Please define the selection rule explicitly so that the choice of C=256 and H=W=32 is reproducible rather than subjective.","section":"Figure 5 caption"},{"comment":"The text says 'we noticed that encoding time as an input coordinate, similar to [3], consistently underperformed FiLM modulation,' but no quantitative comparison is shown for this design decision. A small table or mention in the ablation figure would strengthen this claim.","section":"Section 4, Hyperparameter Selection"},{"comment":"In the Missing-visits scenario, only classical interpolation methods are compared; the absence of DL baselines in this scenario should be stated explicitly in the text, since it currently reads as a broader comparison.","section":"Section 3, Comparison Methods"},{"comment":"The header 'Extrapola tion' contains an unnecessary space; this is a trivial typo but should be corrected.","section":"Table 1 header"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible and interesting application of INRs to a clinically relevant problem, and the authors have made their code available. The main weakness is not the method itself but the evidence quality: a six-eye test set, a duplicated baseline row in Table 1, and an abstract claim that is contradicted by the paper's own results. These issues are fixable within the manuscript's scope, so I would not reject. However, the numerical inconsistency and the lack of statistical support must be resolved before the claims can be taken as established. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What to know: this is the first application of implicit neural representations to modelling geographic atrophy progression, and it does something sensible with them. It couples a FAF reconstruction head and a GA segmentation head on a shared modulated INR, with per-eye latent vectors and time/age conditioning. The two-layer feature concatenation into the segmentation head is a small but real tweak over prior work. The writing is clear, the method is reproducible (code is released), and the authors are refreshingly honest in the conclusion that copy-forward gives the best FAF reconstruction in their short follow-up setting.\n\nThe soft spots are real but addressable. The abstract claims the method delivers the lowest lesion-area MAE and highest DICE \"without sacrificing FAF image quality,\" yet Table 1 shows copy-forward outperforming the method on PSNR, SSIM, and LPIPS in both Scenario 1 and Scenario 2. The conclusion already admits this, so the abstract just needs to be aligned with the evidence. Bigger issue: Table 1 Scenario 2 lists Linear extrap. with exactly the same metrics as Copy-forward (17.86 / 0.61 / 0.14 / 0.86 / 6.98 / 0.38). That is a copy-paste error, and it undermines confidence in the numerical comparison until fixed. Finally, the test set is six eyes. The headline numbers—DICE 0.91 vs 0.86, MAE 0.20 vs 0.38—come with overlapping error bars and no significance test. On six eyes, a paired test would likely leave the superiority claim unproven. That is the biggest scientific weakness, and it is a matter of reporting rather than a flawed method.\n\nThe method itself is a reasonable extension of prior INR progression work to a new medical domain, and the missing-visit interpolation results are the most convincing part of the evaluation. The paper is for ophthalmic image analysis researchers and INR practitioners looking for clinical applications. It deserves a serious referee: the idea is worth publishing, but the claims need to be scaled back, the table fixed, and ideally some uncertainty quantification or a larger test set added. I would not desk-reject it.","headline":"A plausible first INR-based model for geographic atrophy progression, with a sensible joint reconstruction/segmentation design, but the headline comparison rests on 6 test eyes and a duplicated baseline row in Table 1, so the evidence is thinner than the abstract suggests.","tokens_in":9376,"tokens_out":1586,"would_cite":false,"duration_ms":36242,"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":"An implicit neural representation with a per-eye latent vector models geographic atrophy as a continuous trajectory and predicts future FAF images and GA segmentations, reaching DICE 0.91 in test-time adaptation.","keywords":["implicit neural representations","geographic atrophy","age-related macular degeneration","disease progression","fundus autofluorescence","test-time adaptation","image segmentation","longitudinal modelling"],"falsifier":"Take an eye with at least three visits in which the GA lesion grows a new disconnected atrophic island between visits instead of only expanding the existing boundary. Fit the model on the first two visits, hold out the third, and check whether the predicted segmentation contains the island. Because the per-eye latent is shared and the trajectory is continuous, the model can only deform the existing lesion; missing the new island at the held-out visit would falsify the smooth single-trajectory assumption.","tokens_in":8385,"feed_emoji":"👁️","tokens_out":9487,"duration_ms":94032,"temperature":0.7,"pith_summary":"The paper sets out to establish that a single implicit neural representation, conditioned on an eye-specific latent vector plus time and age, can capture the full geographic-atrophy trajectory of an individual eye from a handful of fundus autofluorescence images. If this is right, the same model can generate previously unseen FAF images and GA segmentations at past, intermediate, and future time points, which matters because GA progression is highly individualised and hard to forecast. In the main test-time adaptation scenario the model reports DICE 0.91 and lesion-area MAE 0.20 mm² for future segmentations, and the lowest lesion-area MAE (0.15 mm²) for interpolating a held-out visit among the methods compared. A sympathetic reader would take the paper's claim to be that disease progression in GA is smooth enough in latent space to be learned and extrapolated per patient, even when training data are scarce.","feed_headline":"Neural net forecasts geographic atrophy growth in each eye","feed_subtitle":"Per-eye time-conditioned model predicts future lesion shape and area, scoring DICE 0.91 in test-time adaptation.","key_machinery":"The central object is a time-conditioned implicit neural representation (INR): an auto-decoding multi-layer perceptron with SIREN periodic activations that carries a shared set of weights $\\theta$ for all eyes. Each eye $i$ owns a spatial latent vector $z_i \\in \\mathbb{R}^{C\\times X_1\\times X_2}$, sampled at a coordinate $x$ by bilinear interpolation and concatenated with the time $t$ (weeks since baseline) and the patient's age $\\mathrm{age}_t$; modulation layers map this combined code into per-layer scale and shift, so the same network is steered into each individual's trajectory. The network splits into a reconstruction head $f^{\\mathrm{faf}}_\\theta$ that outputs FAF pixel intensity (MSE loss) and a segmentation head $f^{\\mathrm{seg}}_\\theta$ that consumes the last two hidden layers and outputs a two-class GA probability (DICE plus binary cross-entropy loss). Training optimises the joint loss over all eyes and visits, and test-time adaptation freezes $\\theta$ while optimising only the new eye's latent against the available visits, so a single forward pass generates the requested past, intermediate, or future image and segmentation. This trajectory-in-latent-space mechanism is what carries the paper's claim.","core_discovery":"The paper claims that geographic atrophy (GA) progression in late dry age-related macular degeneration can be modelled as a continuous, patient-specific trajectory rather than as a sequence of independent visit images. Concretely, a shared implicit neural representation, fitted jointly over many eyes, is conditioned by each eye's own latent vector together with the time elapsed since baseline and the patient's age, and is trained to output both the FAF image intensity and the GA segmentation label at any queried coordinate. At test time the shared network is frozen and only the new eye's latent is optimised on the available visits; a single forward pass then yields held-out past, missing, and future images and segmentations. On an in-house longitudinal cohort of 37 eyes the method reports the best segmentation metrics among compared methods in the test-time adaptation scenario (DICE 0.91, Hausdorff distance 6.78 px, lesion-area MAE 0.20 mm²) and the lowest lesion-area MAE for interpolating a missing visit (0.15 mm²), at image reconstruction quality comparable to the baselines. The paper's conclusion is that INR-based trajectory modelling is a viable low-data way to forecast individual lesion shape and area.","pith_inferences":["Beyond the paper, the same latent-trajectory machinery could transfer to other slowly progressing, sparsely sampled diseases such as glaucoma or diabetic retinopathy, provided the smoothness assumption is verified; nothing in the paper tests that transfer.","Beyond the paper, the latent vector space could be probed by interpolating between two eyes' latents to generate intermediate anatomies, a check the paper does not perform but which would directly test whether the representation is a true progression manifold.","Beyond the paper, a comparison on longer follow-up with more visits would likely widen the gap over copy-forward, since the paper itself notes that the short follow-up period makes copy-forward surprisingly strong on image reconstruction metrics."],"forward_implications":["A clinician could fit a new patient's eye from one or two visits and generate the expected GA lesion shape and area at any future or intermediate time, helping to communicate prognosis.","Because the model produces both FAF images and segmentations at arbitrary query times, lesion-area MAE can be read off at the exact time point a trial protocol specifies, without waiting for that visit to occur.","The method works with only 26 training eyes, so INR-based trajectory modelling is a viable route in rare or low-data medical-imaging settings where large annotated longitudinal cohorts do not exist.","The reported best-case numbers (DICE 0.91, MAE 0.20 mm² in Scenario 2; DICE 0.87, MAE 0.15 mm² for missing visits) support using the model to monitor disease evolution at the individual level."],"supporting_citations":[{"why":"It supplies the auto-decoder design with one latent vector per subject and time conditioning, which this paper adapts to geographic atrophy.","marker":"[4]"},{"why":"It provides the SIREN periodic activation function used as the INR backbone.","marker":"[22]"},{"why":"It provides the FiLM-style modulation layers that map latent and time variables into per-layer scale and shift parameters.","marker":"[18]"},{"why":"It provides the spatial 3D latent vectors with bilinear interpolation that let the model encode per-eye anatomy.","marker":"[5]"},{"why":"It provides the idea of coupling a segmentation head to a reconstruction head, which this paper extends by feeding the last two hidden layers into the segmentation head.","marker":"[25]"},{"why":"ImageFlowNet is the main deep-learning baseline for future FAF prediction, and the paper's two prediction scenarios are aligned with its setup.","marker":"[15]"},{"why":"The time-conditional UNet comparison model (T-UNet) is derived from this denoising diffusion model paper.","marker":"[10]"},{"why":"It supplies the image-to-image Schrödinger bridge used to build the T-I2SBUNet comparison model.","marker":"[16]"},{"why":"It documents the in-house longitudinal FAF cohort, with 37 eyes and up to four visits, that supplies the training and test data.","marker":"[24]"},{"why":"RIFT registration aligns follow-up FAF images to baseline, a necessary preprocessing step for per-pixel trajectory learning.","marker":"[14]"}],"fun_headline_variants":["Implicit neural nets map every eye's blindness progression","Per-eye AI forecasts future retinal scar shape and area","Patient-specific neural model predicts geographic atrophy growth","Implicit neural representation forecasts retinal lesion evolution per eye","Neural net predicts blind spot growth years ahead from few scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each eye's geographic atrophy follows one smooth, continuous trajectory that a single low-dimensional latent vector, combined with time and age, can represent; if a lesion grows in jumps, appears in new disconnected places, or the latent cannot separate stable anatomy from visit-specific noise, the predicted segmentations will be inaccurate even when the model fits its training data.","fun_headline_variants_meta":{"raw":{"variants":["Implicit neural nets map every eye's blindness progression","Per-eye AI forecasts future retinal scar shape and area","Patient-specific neural model predicts geographic atrophy growth","Implicit neural representation forecasts retinal lesion evolution per eye","Neural net predicts blind spot growth years ahead from few scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000672,"raw_usage":{"total_tokens":3067,"prompt_tokens":957,"completion_tokens":2110,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2034}},"tokens_in":573,"tokens_out":2110,"duration_ms":15272,"temperature":1.0,"reasoning_tokens":2034,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:59:38.973270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an eye with at least three visits in which the GA lesion grows a new disconnected atrophic island between visits instead of only expanding the existing boundary. Fit the model on the first two visits, hold out the third, and check whether the predicted segmentation contains the island. Because the per-eye latent is shared and the trajectory is continuous, the model can only deform the existing lesion; missing the new island at the held-out visit would falsify the smooth single-trajectory assumption.","supporting_citations":[{"cited_title":"In: Gee, J.C., Alexander, D.C., Hong, J., Iglesias, J.E., Sudre, C.H., Venkataraman, A., Golland, P., Kim, J.H., Park, J","cited_arxiv_id":null,"evidence_quote":"It supplies the auto-decoder design with one latent vector per subject and time conditioning, which this paper adapts to geographic atrophy."},{"cited_title":"In: Advances in Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"It provides the SIREN periodic activation function used as the INR backbone."},{"cited_title":"In: 2021 IEEE/CVF International Conference on Computer Vi- sion (ICCV)","cited_arxiv_id":null,"evidence_quote":"It provides the FiLM-style modulation layers that map latent and time variables into per-layer scale and shift parameters."},{"cited_title":"IEEE Transactions on Medical Imaging45(2), 738–750 (Feb 2026)","cited_arxiv_id":null,"evidence_quote":"It provides the spatial 3D latent vectors with bilinear interpolation that let the model encode per-eye anatomy."},{"cited_title":"In: Gee, J.C., Alexander, D.C., Hong, J., Iglesias, J.E., Sudre, C.H., Venkataraman, A., Golland, P., Kim, J.H., Park, J","cited_arxiv_id":null,"evidence_quote":"It provides the idea of coupling a segmentation head to a reconstruction head, which this paper extends by feeding the last two hidden layers into the segmentation head."},{"cited_title":"In: Modelling GA Progression using INRs 11 ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)","cited_arxiv_id":null,"evidence_quote":"ImageFlowNet is the main deep-learning baseline for future FAF prediction, and the paper's two prediction scenarios are aligned with its setup."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the image-to-image Schrödinger bridge used to build the T-I2SBUNet comparison model."},{"cited_title":"Acta Ophthalmologica (2026)","cited_arxiv_id":null,"evidence_quote":"It documents the in-house longitudinal FAF cohort, with 37 eyes and up to four visits, that supplies the training and test data."},{"cited_title":"IEEE Transactions on Image Processing 29, 3296–3310 (2020)","cited_arxiv_id":null,"evidence_quote":"RIFT registration aligns follow-up FAF images to baseline, a necessary preprocessing step for per-pixel trajectory learning."}],"review_version":1}