acceptodds
Under review as a conference paper at ICLR 2027

Localizing Active Dynamics in Scientific Images: Physics-Grounded Visual Explanations for Segmentation Models

Abstract

Scientific image segmentation identifies structures of interest, but it does not necessarily reveal where the underlying physical or biological mechanism that produced them is currently active. We use active dynamics to denote image regions in which the governing mechanism is currently changing the observed system. These regions differ from the segmentation target. For example, a mask may identify material that has already crystallized, while the remaining transformation occurs in the untransformed regions between crystals. This creates a structural limitation for gradient-based explanation: it attributes the segmentation prediction to supporting image evidence, hence inherently drawn toward the segmented structure, even when the governing dynamics are active in a different region. We introduce PGAD-X (Physics-Grounded Active-Dynamics eXplanation), a framework that produces an active-dynamics explanation alongside the segmentation mask. PGAD-X encodes the governing law and its characteristic spatial scale to condition shared visual features, and uses attention-based explanation to localize where the corresponding dynamics are expressed. Our key design decouples the explanation from the segmentation gradient while grounding both outputs in the same visual representation, enabling the model to localize physically active regions, rather than merely restate the predicted mask. A lightweight second stage aligns the explanation with automatically derived physics references while keeping the segmentation backbone fixed. Across three microscopy datasets, three governing mechanisms, and three segmentation backbones, PGAD-X improves active dynamics localization over gradient-based and latent-alignment baselines while preserving or improving segmentation accuracy. We release the physics references and evaluation protocol.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI, Physics-Guided Learning, AI for Science, Scientific Imaging, Interpretable Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sVCxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PGAD-X, based on scientific image segmentation, adds an active dynamic heatmap. The Physical Prefix Encoder (PPE) encodes the description of control laws and scale information, while the Cross Attention Explanation Head (CAEH) generates the heatmap from the shared visual features. The second-stage training aligns this heatmap with the reference image p(x) generated by the rules. The experiments cover three microscopic image datasets and three backbone networks, among which AFM also uses independent expert references and time validation.

Critical strengths
  1. Separating the segmented structures from the changing areas is of practical value. Limiting the evaluation to the region where the mechanism is likely to occur is particularly commendable: Table 3 shows that the overall score of the full image may reward the ability to distinguish the foreground from the background, rather than the ability to identify the target activity.

  2. AFM provides independent expert reference and time validation. In Table 13, the improvement of PGAD-X over the auxiliary head on SegFormer and SAM2 provides some evidence that goes beyond the fitted training proxy.

Critical weaknesses
  1. In Table 6, after using non-scientific text and re-training, the AUC of AFM/SegFormer-B5 was 0.829, the correct description was 0.824, and the additional benefit of accurately describing the correct mechanism was not reflected. The partial time prediction improvement in Table 13 is commendable, but it still needs text ablation to confirm whether it comes from the mechanism semantics.

  2. BBBC039 marks the cell nuclei in the top 25% of the single-frame CV ranking as positive, but "high CV" does not equal "CV is increasing". The evaluation of the two biological datasets still follows the training rules, which is insufficient to verify true proliferation or division. Therefore, the relevant explanations in Appendix D should be supplemented with independent evidence or the conclusion should be narrowed.

Questions for authors
  1. Can the paired random seeds be used to repeat the text ablation experiment in Table 6, reporting the differences and their uncertainties? Applying the same ablation to Table 13 will also help to verify whether the improvement in time prediction depends on the correct description.

  2. What is the basis for Appendix D to interpret the heat map of BBBC039 as a dividing cell nucleus or mitotic index, and to interpret the heat map of LIVECell as proliferation activity? Can these claims be narrowed down to the evaluated proxy targets?

  3. D1 requires shared representation, and CAEH meets this requirement. Should the "model fidelity" mentioned throughout the text be understood only in this sense? Formulas 2 and localization scores cannot yet prove the impact of heatmap responses on model decisions.

  4. Table 4 has already compared different reading methods at the same first-stage checkpoint. For the complete model comparison in Table 1, if the image intensity scaling is unified and the GradCAM at the second-stage checkpoint is supplemented, what changes will occur in the results of BBBC039? The relevant preprocessing differences can be found in Appendix C.4.

2 comments

W1/Q1

Thank you. We repeated the ablation with five paired seeds (AFM, SegFormer-B5; Appendix G, Table 14); differences to PGAD-X with 95% CIs:

ChangeExpert AUCTemporal AUC
Other scientific text+0.002 [−0.000, 0.004]+0.002 [−0.030, 0.035]
Non-scientific text+0.003 [−0.000, 0.007]+0.004 [−0.028, 0.037]
Flat scale prior−0.006 [−0.009, −0.003] (p = 0.007)+0.009 [−0.013, 0.031]

Neither the static map nor the temporal gain depends on the wording, whereas what the hypothesis specifies does matter: the scale prior, and the hypothesis itself, since a model trained under a different hypothesis follows that hypothesis (diagonal vs. off-diagonal AUC 0.79/0.55 on AFM; Appendix H). We now state this explicitly: the hypothesis enters PGAD-X through its computable terms (reference map, characteristic length, prior type), which also makes the map robust to how an expert or an AI system phrases it.

W2/Q2

The readings in Appendix D rested on the training rule alone, so we added independent evidence and narrowed the claims. On the Cell Tracking Challenge HeLa time-lapse with lineage ground truth (305 divisions within 30 min among 33,357 nuclei; Appendix J), high intra-nuclear CV predicts imminent division (AUC 0.824 at k = 1, 0.600 at k = 4), whereas nuclear area and DNA content predict it inversely (0.254, 0.262). In Appendix M (trained on sequence 01, tested on 02), PGAD-X ranks the nuclei that divide next at anticipation AUC 0.747/0.766/0.720, above PGAD-X-Aux (0.662/0.756/0.619); BBBC039 models applied to HeLa without retraining reach 0.758–0.800. For LIVECell, the "exponential-phase / doubling-time" reading is removed and the map is described as where the stated hypothesis places proliferation; Appendix D now describes the BBBC039 map as the nuclei that the condensation hypothesis selects, citing the HeLa evidence.

Q3

Yes. We renamed the requirement "shared representation" (now D3, Sec. 2.2): H_vis is derived from the same features that produce H_seg, so aligning the explanation acts on, and should preserve, the segmentation, which we treat as a trade-off and evaluate (S1 vs. S2 IoU in Table 1). It is not a claim that H_vis is the causal basis of the segmentation, and the interventions in Appendix I are reported as tests of how the map responds to the input: perturbing the map's own top region changes it more than a random perturbation, on LIVECell this region barely affects the segmentation while the masked region does, and re-initializing the trained weights (Adebayo et al., 2018) drops its AUC to chance (mean 0.515).

Q4

All BBBC039 maps in Table 1 are now computed from the same percentile-scaled input that Stage 2 uses (Appendix L); under it the Stage-1 checkpoint segments at IoU 0.612/0.303/0.514 (0.859/0.879/0.684 under the 16-bit scaling it was trained with). BBBC039 AUC:

SegFormerSAM2DINOv2
GradCAM, Stage-1 checkpoint.516.438.538
GradCAM, Stage-2 checkpoint.586.534.574
PGAD-X, same Stage-2 checkpoint.783.729.763

On the same Stage-2 checkpoint PGAD-X is higher on 66/66, 66/66 and 64/66 test images (p < 1e-11), and the same holds on AFM against the expert annotation (GradCAM at Stage 2: 0.582/0.687/0.669; PGAD-X: 0.840/0.835/0.833). Stage-2 training therefore does not by itself make a gradient map physics-aligned.

Reviewer_sVCxOpenReview ↗

Thank you for the author's reply. I'll keep my score.

Thank you for the author's reply. I'll keep my score.

Reviewer_HvbUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines the problem of localizing active dynamics in an image, in contrast to a segmentation problem which is focused on separating out the static structures in an image. As specific problem instances, the paper evaluates on three modalities - the crystalline phase in atomic-force images, cell proliferation of microglia, and chromatin-condensation in nuclei.

The paper proposes a Physics-Grounded Visual Explanation framework, consists of a two-stage model to augment a standard segmentation model. The framework takes as input a governing law specifying the active dynamics, with a specific instantiation as a grounding operator that maps from image information to spatial activity. The overall system consists of a shared feature trunk, a physics prefix encoder that represents the governing law as well as a characteristic-scale prior, and a cross-attention explanation head that connects the two prior components.

When evaluated on the above three data domains, the proposed method shows improvement over baselines of GradCam and the single-stage version of the framework that omits the physics/active-dynamics pathway.

Critical strengths

The problem formulation is interesting and well-motivated. The method improves over the baselines considered, and this holds across the three data domains tested.

Critical weaknesses

A potential fundamental concern is evaluation against a simple baseline of directly predicting the active dynamics without a joint/aligned segmentation model. This comparison is not given in the paper, but seems to be considered in Appendix C.4 - Auxiliary head and trainable scope. Against the simple baseline of directly predicting the supervision signal from frozen stage 1 features, without physics prefix, scale prior, and FiLM conditioning, this auxiliary head baseline seems to outperform the proposed model on the datasets that were tested for this part of the study (BBBC039 and AFM). A separate validation study is then proposed, but here there appears to be no firm conclusion as far as which method works better, between GradCAM, auxiliary head, and proposed PGAD-X. This seems to call into question the need for the proposed framework when this simple alternative seems to work equally well if not better.

Questions for authors

Is there a clear case in which PGAD-X outperforms the auxiliary head baseline and demonstrates the utility of the additional proposed modules?

1 comment

W1/Q1

Thank you. Yes, in two respects. We now include the auxiliary head in Table 1 as a variant of our method, PGAD-X-Aux. On the static references the two are comparable (within 0.007 AUC on six of nine pairs), which shows that our Stage-2 rank objective is effective on its own. First, the variant reads frozen Stage-1 features and cannot change the segmentation, whereas PGAD-X conditions the shared features and raises IoU on seven of nine pairs (up to +0.048). Second, in Appendix M we score both on three time-lapse datasets, asking whether the map at time t ranks the units that change next, with labels from later frames or lineage that neither was trained on (anticipation AUC, PGAD-X vs. PGAD-X-Aux):

SegFormerSAM2DINOv2
AFM (crystallizes).876 vs .836.871 vs .855.876 vs .786
LIVECell (newly covered).659 vs .646.678 vs .641.641 vs .619
HeLa (divides).747 vs .662.766 vs .756.720 vs .619

PGAD-X is higher on all nine pairs, significantly at both horizons on eight (Table 13). The earlier ambiguity in the validation study came from DINOv2; under the configuration selected on validation, PGAD-X also leads with DINOv2 (+0.090 on AFM).

Reviewer_vHK5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces "active-dynamics explanation" for scientific image segmentation: instead of explaining which pixels support the mask, the model should localize where the governing physical or biological law is currently active. In the motivating AFM example, crystallization continues in the untransformed corridors between crystals, not in the crystallized phase that the mask marks; the authors report a mean correlation of −0.10 between the mask and the expert-annotated active region, which supports that the two targets are genuinely different. The method (PGAD-X) has three parts: a shared feature trunk (backbone plus feature pyramid, with fusion weights initialized from a physics-based scale prior); a physics prefix encoder that turns a description of the law (CLIP text of the law, a length-scale prior, a boundary/density type flag) into 8 prefix tokens that condition the backbone through FiLM and low-rank adapters; and a cross-attention explanation head whose response magnitude is the explanation map, computed without backpropagating through the segmentation loss. Stage 1 trains segmentation; Stage 2 freezes the backbone and aligns the map to a deterministic reference p(x) with a differentiable Spearman rank loss, restricted to an "admissible domain" in which the mechanism could be active. Evaluation covers three microscopy datasets (AFM crystallization, 46 frames; LIVECell BV2 microglia; BBBC039 nuclei) and three backbones (SegFormer-B5, SAM2 Hiera-L, DINOv2). Segmentation IoU moves by between −0.008 and +0.028 after Stage 2, and explanation AUC improves over GradCAM and a JEPA-style latent alignment baseline in all nine dataset/backbone combinations (e.g., AFM AUC 0.824–0.837 for PGAD-X versus 0.648–0.722 for GradCAM).

Critical strengths
  1. The problem is real and the motivation is well made. The difference between what the mask marks and where the process is active is quantified rather than just asserted (mask versus expert reference correlation −0.10), and the three requirements (model faithfulness, physics alignment, complementarity to the mask) are stated cleanly. I am not aware of earlier work that formulates this exact problem for scientific segmentation.
  2. The evaluation protocol is useful on its own. Restricting scores to the admissible domain, using threshold-free ranking metrics, and reporting trivial floors (raw intensity, inverted intensity, mask, boundary distance) all address a real weakness in how explanations are usually scored. Table 3 is the clearest example: on BBBC039, copying the segmentation mask scores 0.884 AUC on the full frame but exactly 0.500 inside the nuclei, which shows why the domain restriction is needed.
  3. The paper is honest about where its own evidence is weak, more so than most submissions. The authors state that BBBC039 is evaluated against the same rule used for training (p. 18), that Stage 2 also acts as additional segmentation training (the λ=0 run already lifts the explanation readout to 0.562), and that the temporal test is only a partial validation because a boundary-distance baseline wins.
  4. Reporting is thorough. The AFM reference comes from 22 materials scientists and is independent of the training target (Table 15: the morphological closing recovers 0.981 recall at 5 px against the expert annotations, while a watershed reconstruction based only on mask geometry recovers almost nothing). Three-seed variance is at most 0.004 in AUC, the closing-radius sensitivity is small, and the cost of the extra head is analyzed.
Critical weaknesses
  1. My main concern is that the physics conditioning does not clearly do the work the paper credits it with. A matched control, a three-layer convolutional head trained on the same rank objective with no physics prefix, no scale prior, and no FiLM, matches or beats PGAD-X: it wins on all three backbones for BBBC039 (AUC 0.785, 0.784, 0.770 against 0.758, 0.712, 0.748), and on the independent AFM reference the paper itself calls the two "on par" with all alignment frames (0.833 versus 0.824), with the auxiliary head ahead by 0.02–0.03 when fewer frames are used (Table 12). The descriptor ablation is harder to explain: replacing the law text with unrelated non-scientific text gives 0.829 AUC on AFM, marginally above the 0.824 for the real law text (Table 6). The only remaining advantage of the full model is in the temporal test on two of three backbones, and there a boundary-distance baseline (0.939/0.919) still beats all learned maps. As it stands, the evidence better supports the claim that rank supervision on a mask-derived reference moves a learnable readout away from the mask. I understand this is not what the title and the contributions claim, so either the scope needs to change, or the ablations need to show a regime where the physics ingredients win.
  2. The supervision signal is derived from the mask. All three grounding operators are deterministic functions of the image and the segmentation mask; for AFM the paper describes p(x) as a morphological closing of the segmentation mask. The model is therefore being asked to reproduce a hand-crafted mask function, which is exactly what the auxiliary-head control does. The practical bottleneck is then the grounding operator itself, which has to be designed per dataset with domain expertise. This puts the "plug-able governing laws" claim (p. 4) in some tension with the experimental setup, and the paper does not quantify how often such an operator can be written without deep domain knowledge.
  3. The statistics are thin. AFM is one time series of one field of view (46 frames, 14 used for evaluation); LIVECell has 39 images and BBBC039 has 197. No confidence intervals or significance tests are reported for the main comparisons, and the gap between PGAD-X and the auxiliary head on AFM (0.833 versus 0.823 ± 0.004, Tables 12 and 14) is roughly 2.5 seed standard deviations. The "all nine pairs" claim in the abstract also includes the BBBC039 results, which the authors themselves flag as sharing the supervision rule.
  4. A couple of comparisons I would expect by now. On the baselines side, at least one microscopy foundation-model baseline (μSAM, Nature Methods 2025) on LIVECell and BBBC039. On the faithfulness side, an interventional test of the explanation maps (perturb the active region versus the masked region) rather than only agreement with a reference; this matters because requirement D1 is satisfied by construction (shared features), and the recent faithfulness literature suggests attention-based readouts do not reliably reflect the model's decision process. Separately, "physics-admissible domain" is introduced as a new concept but overlaps with existing terminology from both XAI (admissible explanation spaces) and PINN/surrogate models (physical admissibility); the paper should disambiguate against both.
  5. The bibliography is in order except for three entries with errors. Johnson (1939) is missing the co-author R. F. Mehl, and the journal name is written twice ("Transactions of Transactions of…"). The LoRA entry spells "Liang Wang" where the arXiv record (2106.09685) has Lu Wang, and the venue string "Iclr, 1(2):3" is malformed. Nunes et al. 2026 is now in print as EMBO J 45(15):5357–5387, so the "pp. 1–31" placeholder should be updated.
Questions for authors
  1. Tables 6, 11, and 12 suggest the physics ingredients are not needed once alignment is in place, while Table 7 suggests they matter before it. Could the authors characterize this boundary more systematically, for instance AUC as a function of the number of alignment frames, with PGAD-X and the auxiliary head side by side? If the descriptor's role turns out to be limited to initialization, would they re-scope the contribution as rank-supervised localization of mask-external activity, with the physics descriptor as a zero-shot initialization?
  2. Could the authors add interventional faithfulness tests (perturbing the active region versus the masked region) and a μSAM-class baseline on LIVECell and BBBC039?
  3. Could confidence intervals or significance tests be reported for the AFM comparisons, and the BBBC039 results presented separately given that they share the supervision rule? If feasible, could the temporal validation be extended beyond a single time series?
  4. For a new dataset where a domain expert cannot construct a plausible grounding operator, what does the method fall back to, and how does the explanation quality degrade?
2 comments

W1/Q1

Thank you. We now report the auxiliary head as a variant of our method, PGAD-X-Aux (Sec. 4 Exp-1; Appendix M). The variant is comparable to PGAD-X on the static references of Table 1 (within 0.007 AUC on six of nine pairs), which confirms that the Stage-2 rank objective is effective on its own. Conditioning the shared features adds two things the variant cannot provide, segmentation that Stage 2 can improve and anticipation of the dynamics on three time-lapse datasets with labels from later frames or lineage:

PGAD-X − PGAD-X-AuxSegFormerSAM2DINOv2
IoU, AFM / LIVECell / BBBC039+.015 / −.007 / +.010+.007 / +.001 / −.002+.014 / +.048 / +.002
Anticipation AUC, AFM+.040+.016+.090
Anticipation AUC, LIVECell+.013+.037+.022
Anticipation AUC, HeLa+.085+.010+.101

The anticipation gaps are significant at both horizons except HeLa/SAM2 (Table 13). The boundary-distance baseline is now reported under the same protocol (Appendix M): it reaches 0.936 on AFM and 0.725 on LIVECell, where growth proceeds from existing structures, but both variants receive this information equally, it does not localize the hypothesized activity (expert AUC 0.740, AP 0.066 vs. 0.839, 0.194 for PGAD-X; Table 2), and on HeLa, where division cannot be read from distance to the mask, PGAD-X leads by up to 0.10. The paired-seed ablation (Appendix G) shows that the map follows what the hypothesis specifies rather than its wording: rewording changes expert AUC by at most 0.004, a flat scale prior lowers it (p = 0.007), and training under another hypothesis moves the map to it (diagonal vs. off-diagonal 0.79/0.55, Appendix H).

On Q1, we ran the curve (Appendix N, Table 20): PGAD-X and PGAD-X-Aux on AFM with 0, 2, 4, 8, 16 or all alignment frames, three seeds each. Both reach the same expert AUC as alignment data grow, while PGAD-X leads the temporal test at most budgets (temporal AUC, PGAD-X − PGAD-X-Aux):

2 frames4816all
SegFormer+.039+.006+.058−.034+.103
SAM2+.051+.035+.024+.128+.020
DINOv2+.015+.053−.005−.037+.136

The descriptor therefore acts through the conditioning pathway trained in Stage 2 rather than as a zero-shot initialization. We keep the problem definition and describe the contribution as both parts together, rank supervision and the conditioning of the shared features, with PGAD-X-Aux separating the two (Sec. 4).

W2/Q4

The operator only supplies training targets; evaluation uses the expert annotation on AFM and, in Appendix M, later frames or lineage, which no operator produces. Because G_d may now be a hypothesis (Sec. 2.1), an operator need not be correct to be useful: Appendix H gives 2–4 operators of one to three lines per dataset, each model follows the hypothesis it is trained on, and independent data reject the wrong ones (AFM nucleation-dominated growth: expert AUC 0.235; HeLa nuclear area and DNA content: 0.254 and 0.262, against 0.824 for chromatin CV; Appendices H, J). For the fallback, we replaced the designed AFM operator with a generic boundary operator that needs no domain knowledge, and with a deliberately wrong one (expert AUC, Appendix H):

AFM, expert AUCSegFormerSAM2DINOv2
PGAD-X, designed operator.839.827.836
PGAD-X, generic boundary operator.815.834.811
PGAD-X, wrong operator.597.778.525
PGAD-X-Aux, wrong operator.430.598.402

A generic operator keeps most of the quality, and a poor operator is visible against independent evidence.

W3/Q3

All trained methods now use three seeds with paired Wilcoxon tests (Appendix F); Table 13 gives every PGAD-X − PGAD-X-Aux difference on the time-lapse data with its p-value per horizon, and Appendix G gives paired-seed confidence intervals. On AFM the two are within 0.007 on the expert reference (Table 1); with five paired seeds the variant is 0.009 higher there, while PGAD-X is 0.053 and 0.077 higher on the temporal tests (Table 14). BBBC039 stays in Table 1 as in the submission, and its rule is now tested independently: on the Cell Tracking Challenge HeLa time-lapse, lineage shows that high intra-nuclear CV precedes division (AUC 0.824; Appendix J), and PGAD-X trained with this rule ranks the nuclei that divide next at 0.747/0.766/0.720 (Appendix M). Temporal validation now covers three time-lapses (AFM, 46 frames; LIVECell BV2, 32 fields at 4 h steps with 16 held-out test fields; HeLa, two sequences at 30 min steps). For the LIVECell time-lapse, we selected among three hypotheses by their rule alone on validation fields before training any model (free-space proliferation, 0.598 vs. 0.515 for logistic and 0.529 for colony-edge growth; 0.605 on the test fields; Appendix K), which illustrates how the framework tests a hypothesis against independent data.

W4/Q2

We ran both requested interventions and a μSAM baseline (Appendix I, Table 17). First, on the Stage-2 checkpoints we blur 20% of the input pixels taken from the top of H_vis (active), at random inside the predicted mask (masked), or uniformly (random), and record the IoU drop and the rank correlation of H_vis before and after (lower ρ: larger change of the map):

20% of pixelsIoU drop: active / masked / randomρ(H_vis): active / masked / random
AFM, SegFormer.156 / .221 / .183.65 / .87 / .90
AFM, DINOv2.085 / .034 / .013.86 / .95 / .93
LIVECell, SegFormer.069 / .646 / .018.67 / .63 / .95
LIVECell, DINOv2.019 / .198 / .000.79 / .74 / .93
BBBC039, SegFormer.130 / .052 / .044.83 / .74 / .83
BBBC039, DINOv2.140 / .062 / .034.56 / .79 / .88

The map changes more under perturbation of its own top region than under a random one in five of six cases. On LIVECell, where the hypothesis places the activity in the free space around the cells, the two outputs rest on different evidence: the active region barely affects the segmentation, the masked region strongly does. On BBBC039 the active nuclei are themselves segmented objects, so the two perturbations overlap. Second, removing single crystallites by inpainting (38 edits) removes them from the segmentation (0.72 → 0.03), while the PGAD-X map far from the edit changes by only 0.004, against 0.014 for PGAD-X-Aux and 0.034 for GradCAM, and its AUC against the recomputed reference rises from 0.828 to 0.860. Third, re-initializing the trained weights layer by layer (the model-randomization test of Adebayo et al., 2018; 36 models) decorrelates the map (mean |ρ| 0.19) and drops its AUC to chance (mean 0.515). We agree that attention is not by itself a causal account of the decision; requirement D1 of the submission is renamed "shared representation" (now D3) and stated in that sense only (Sec. 2.2).

μSAM (vit_b_lm, no fine-tuning) is evaluated in Appendix F on the time-lapse data, including HeLa, whose nuclei use the BBBC039 chromatin-condensation rule, and on BBBC039. Applying the rule to its segmentation gives static AUC 0.621/0.420/0.581 (AFM/LIVECell/HeLa), as its segmentation of these images is weak (IoU 0.199/0.071/0.470); its features under the same head give 0.811/0.848/0.836 with anticipation AUC 0.853/0.650/0.660, below PGAD-X with SegFormer on static AUC, AP and anticipation AUC on all three datasets. On BBBC039 itself (Table 1 protocol), μSAM segments at IoU 0.599 (PGAD-X: 0.869/0.877/0.686), the rule on its segmentation gives AUC 0.577 and AP 0.319, and its features under the same head give AUC 0.771 and AP 0.532, within the range of PGAD-X across backbones (0.732–0.782) and below PGAD-X with SegFormer (0.782, AP 0.553).

W4 (terminology)

We added this to Sec. 2.2: Ω_adm is a set of image locations used for scoring. It is not an admissible explanation space (XAI) and not physical admissibility of predicted fields (PINNs).

W5

Fixed: Johnson & Mehl (1939), with the duplicated journal name removed; LoRA (Lu Wang; ICLR 2022); Nunes et al., EMBO J 45(15):5357–5387.

Comments on the paper

1 comment

We thank the reviewers for the careful reading; new experiments in the PDF are marked in blue. All three reviews ask what the physics conditioning adds over an auxiliary head trained on the same reference. We now include this head in Table 1 as a variant of our method, PGAD-X-Aux. It is comparable to PGAD-X on the static references, which confirms that our Stage-2 rank objective is effective on its own. Conditioning the shared features adds two things the variant cannot provide: Stage 2 preserves or improves segmentation (IoU raised on seven of nine pairs, by up to 0.048), and PGAD-X anticipates what changes next better on all nine dataset–backbone pairs of three time-lapse datasets, with labels from later frames or lineage (Appendix M). We also renamed "model faithfulness" to "shared representation" (D3), allow G_d to be a hypothesis tested against independent evidence (Sec. 2.1), narrowed the biological readings (Appendix D), and added the requested statistics and ablations (Appendices F–N).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.