{"id":"9652306a-7743-4b75-b3a3-2d5db8b33cba","arxiv_id":"2607.04912","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A time-aware GNN with three asymmetric trajectory losses outperforms vision and SSL baselines for pCR prediction from multi-timepoint DCE-MRI on 585 ISPY-2 patients.","lead":"A graph neural network plus three self-supervised losses on longitudinal DCE-MRI predicts breast-cancer response to chemotherapy better than CNN, LSTM, and SSL baselines on ISPY-2. The work offers a public benchmark and code path for early treatment-response modeling from serial scans.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Asymmetric losses apply only to responders; non-responders receive only a repel term, so the claimed 'trajectory representations' are not symmetrically learned and may not generalize beyond ISPY-2 class balance.","rationale":"The reader correctly flags that the three losses may produce ISPY-2-specific artifacts rather than clinically meaningful trajectories, and that this premise is load-bearing. The more precise soft spot is the explicit one-sided application of those losses: only responders receive the full suite of trajectory constraints, while non-responders are merely repelled. This design choice is never ablated against a symmetric alternative, so the empirical gains in Tables 1–2 cannot yet be attributed to general longitudinal modeling. The concern does not overturn the multi-fold, multi-baseline evidence already presented, nor does it introduce an internal contradiction; it simply tightens the weakest assumption. Hence the verdict remains CONDITIONAL (code still unreleased, single-cohort, absolute metrics moderate) and confidence stays moderate. The concrete re-training experiment would settle whether the asymmetry is essential or incidental.","tokens_in":9687,"tokens_out":599,"duration_ms":5525,"concrete_test":"Re-train the identical architecture and ResNet18 encoder under a fully symmetric loss that applies L_align + L_decorrelate + L_temporal to both responders and non-responders (or, alternatively, only to non-responders) while keeping the same 5-fold splits and hyperparameters. If mean bACC/AUC/MCC drop by more than ~0.02 relative to the reported GNN-pCR numbers, the asymmetry itself—not the graph topology or the temporal objectives—is load-bearing and the clinical-trajectory interpretation weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on three novel complementary self-supervised trajectory objectives that produce clinically meaningful longitudinal representations (Abstract, §2.2, Discussion). Inspection of the loss definitions shows the asymmetry is stronger than the reader noted: L_align (Eq. 1), L_decorrelate (Eq. 2) and L_temporal (Eq. 3) are all gated by the indicator 1{i∈R}; non-responders receive only the repel term of L_align (final L_total formula). Consequently the model never enforces complementary or additive temporal structure on the majority class (381/585 non-responders). The ablation in Table 1 removes each term but never tests a symmetric counterpart that applies the same three objectives to both classes (or to non-responders alone). If the performance lift is driven primarily by this one-sided regularizer rather than by genuine trajectory modeling, the claim that the method captures 'clinically meaningful longitudinal' dynamics is overstated for the full population and for early-response settings (Table 2) where class imbalance is unchanged.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes GNN-pCR, a 3D spatio-temporal framework for predicting pathological complete response (pCR) to neoadjuvant chemotherapy from longitudinal DCE-MRI. Image volumes at each timepoint are encoded by a shared ResNet18 into embeddings that form a directed acyclic graph (edges from every earlier to every later timepoint); GraphSAGE aggregates the graph into a patient-level embedding. Training is driven by an Asymmetric Treatment Response Loss whose three novel terms (population-level responder alignment, adjacent-timepoint decorrelation, and additive temporal consistency of latent differences) are applied only to responders, while non-responders receive only a repel term. On 585 ISPY-2 patients with four timepoints the method reports 5-fold means of bACC 0.6844, AUC 0.7203 and MCC 0.3561, outperforming CNN, CNN+LSTM, DINOv3 and two SSL baselines; ablations, early-timepoint experiments and inter-scan time-difference tests are also provided. Code and a PyPI curation library are promised.","tokens_in":10069,"tokens_out":1090,"duration_ms":10256,"significance":"If the comparative gains hold under external validation, the work supplies a concrete, reproducible imaging-only benchmark for longitudinal pCR prediction and a usable graph-plus-asymmetric-SSL recipe that improves on both supervised sequence models and recent SSL baselines on the same public cohort. The multi-split protocol, shared encoder, Wilcoxon tests, component ablations and planned open-source release are genuine strengths that raise the bar relative to many prior single-split ISPY-2 studies. The clinical claim that the learned representations capture 'clinically meaningful longitudinal' dynamics remains provisional until the asymmetry of the losses and the lack of multi-center data are addressed, but the engineering contribution is already useful to the community.","major_comments":[{"comment":"§2.2, Eqs. (1)–(3) and the final L_total formula: L_align, L_decorrelate and L_temporal are all gated by 1{i∈R}; non-responders (381/585) receive only the repel term. The manuscript never reports a symmetric counterpart that applies the same three trajectory objectives to both classes (or to non-responders alone). Without that control it is impossible to know whether the gains in Tables 1–2 arise from genuine trajectory modeling or from a one-sided regularizer that simply densifies the minority class. This is load-bearing for the claim of 'clinically meaningful longitudinal' representations for the full population.","section":null},{"comment":"All results (Tables 1–2) are obtained on a single public multi-center trial (ISPY-2) with a near-homogeneous acquisition schedule. The Discussion itself notes that inter-scan time differences add little value for this reason. External validation on an independent cohort (or at least a leave-one-site-out analysis) is required before the method can be presented as a generalizable pCR benchmark; otherwise the reported margins may be dataset-specific.","section":null},{"comment":"Table 1 ablation 'w/o GNN' replaces GraphSAGE by a linear head and still yields bACC 0.6610 (close to the full model). Combined with the fact that the three losses already operate on the same embeddings, it remains unclear how much of the lift is attributable to relational message passing versus the asymmetric SSL objectives alone. A clearer isolation (e.g., GNN trained only with supervised loss, or losses applied after mean-pooling) would strengthen the architectural claim.","section":null}],"minor_comments":[{"comment":"Abstract and p. 8: 'medical imagaging' is a typographical error; correct to 'imaging'.","section":null},{"comment":"§3 Implementation Details: DINOv3 is applied to a single 2D axial slice while all other methods use full 3D volumes; the comparison is therefore not modality-matched and should be flagged more explicitly.","section":null},{"comment":"Eq. (1): the random permutation π is described only in text; a short formal definition would improve reproducibility.","section":null},{"comment":"Table 2 caption and body: the inter-scan time-difference experiment is interesting but the near-null result for most methods is under-discussed relative to the claim that temporal modeling is clinically meaningful.","section":null},{"comment":"References [23] and [24] are arXiv preprints; if they remain unpublished at camera-ready, note the version and access date.","section":null}],"recommendation":"major_revision","confidential_remarks":"The technical novelty is real but incremental; the main risk is over-claiming clinical generality from a single-trial, asymmetric-loss design. If the authors supply the symmetric-loss ablation and at least one external or site-held-out experiment, the paper becomes a solid methods contribution suitable for the journal. Without those, it is closer to a well-executed ISPY-2 engineering note."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is straightforward: on 585 ISPY-2 patients with four DCE-MRI time points they get mean 5-fold bACC 0.6844 / AUC 0.7203 / MCC 0.3561 with a GraphSAGE head on a strict temporal DAG of ResNet18 embeddings, plus three self-supervised losses that are applied only to responders (population attract, adjacent-time decorrelation, additive latent-difference consistency) while non-responders only get a repel term. That package beats the vision and SSL baselines they re-ran under identical splits and encoder, and the ablations show each piece contributes.\n\nWhat is actually new is the combination, not the individual ingredients. GraphSAGE, ResNet features, and contrastive-style alignment are standard; the strict forward-only DAG topology plus the deliberate one-sided trajectory regularizers are not. They also give a clean multi-fold benchmark, early-timepoint curves, and a quick check on inter-scan deltas (mostly null, as expected from the trial design). Code and a PyPI curation library are promised; that is real value if they ship.\n\nThe soft spots are real but proportionate. Absolute performance is still moderate for a clinical surrogate. Everything is single-cohort ISPY-2; no external multi-center set. The stress-test note is correct that the three trajectory losses never touch the majority non-responder class (381/585), so claims of “clinically meaningful longitudinal” modeling for the full population rest on an asymmetric design that is never compared to a symmetric counterpart. That is a design choice they justify (heterogeneous non-responders), not a hidden bug, but it does leave the generalization story thinner than the abstract implies. Free parameters (SGD schedule, embedding dim, fixed T=4) are ordinary for this literature.\n\nThis is for people who already work on longitudinal medical imaging or pCR prediction and want a reproducible baseline plus a concrete GNN+SSL recipe. It is not a clinical decision tool yet. I would send it to referees; the experiments are careful enough and the contribution is concrete enough to deserve a proper review rather than a desk reject. Engage if you care about trajectory SSL on serial MRI; skip if you need multi-center numbers or symmetric class modeling.","headline":"Solid empirical package on ISPY-2: temporal DAG + GraphSAGE plus three asymmetric trajectory losses that beat re-implemented baselines; absolute numbers stay modest and the asymmetry is intentional but under-tested.","tokens_in":10674,"tokens_out":561,"would_cite":true,"duration_ms":5078,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A graph model of DCE-MRI trajectories predicts breast cancer complete response better than vision and self-supervised baselines.","keywords":["GNNs","Longitudinal","Self-Supervised","Breast Cancer","pCR prediction","DCE-MRI","treatment trajectories","GraphSAGE"],"falsifier":"Retrain and re-evaluate the identical GNN-pCR pipeline on an external multi-center DCE-MRI cohort whose inter-scan intervals vary substantially; if balanced accuracy and AUC fall to the level of the CNN or LSTM baselines, the claim that the trajectory losses capture generalizable longitudinal dynamics is falsified.","tokens_in":10613,"feed_emoji":"🩺","tokens_out":913,"duration_ms":7294,"temperature":0.7,"pith_summary":"The paper argues that pathological complete response after neoadjuvant chemotherapy can be predicted more accurately by treating each patient's longitudinal DCE-MRI scans as a directed acyclic graph of time-ordered features and learning those trajectories with three complementary self-supervised objectives. On 585 patients from the public ISPY-2 cohort, each with four scans, the resulting GNN-pCR model improves balanced accuracy, AUC and Matthews correlation over CNN, LSTM and existing longitudinal self-supervised methods under identical 5-fold splits. Ablations show that population-level alignment of responders, decorrelation of successive time points, and additive temporal consistency each contribute, and that the same gains appear even when only early time points are available. The work therefore claims that explicit relational modeling of treatment trajectories, rather than simple concatenation or recurrence, better captures the imaging dynamics that distinguish responders from non-responders.","feed_headline":"Graph of MRI trajectories beats baselines on pCR prediction","feed_subtitle":"Three self-supervised losses plus a temporal DAG lift accuracy on 585 ISPY-2 patients","key_machinery":"Asymmetric Treatment Response Loss: three complementary self-supervised terms (population-level alignment that attracts responders while repelling non-responders, patient-level decorrelation of successive embeddings, and temporal consistency that requires the final embedding to match the sum of intermediate latent differences) applied only to responders, together with a temporal DAG whose edges run strictly forward in time.","core_discovery":"A time-aware GraphSAGE network operating on a directed acyclic graph of ResNet embeddings, trained with three asymmetric trajectory losses (responder alignment, adjacent-timepoint decorrelation, and additive temporal consistency), yields higher 5-fold pCR prediction performance (bACC 0.6844, AUC 0.7203, MCC 0.3561) than vision and self-supervised baselines on 585 ISPY-2 patients, establishing that structured longitudinal graph modeling improves individualized treatment-response prediction.","pith_inferences":["Because the alignment loss deliberately omits an attract term for non-responders, the method implicitly treats non-response as a heterogeneous residual class; this design choice may limit calibration when partial responders form a clinically important intermediate group.","The same DAG-plus-asymmetric-loss construction could be tested on longitudinal CT or PET trajectories in other solid tumors where complete-response surrogates exist.","If the additive temporal-consistency term is the main driver of early-prediction gains, replacing it with a learned ODE or continuous-time GNN layer would be a direct next experiment."],"forward_implications":["Early-response models that use only the first two time points can still outperform concatenation and LSTM baselines, supporting earlier therapy adaptation.","The same graph-plus-asymmetric-loss recipe can be reused for other longitudinal imaging prediction tasks beyond breast-cancer pCR.","Public release of the curation library and code enables standardized multi-split benchmarks for future longitudinal pCR methods.","Including raw inter-scan time gaps yields only modest gains under ISPY-2's near-uniform schedule, suggesting the graph topology already encodes the dominant temporal structure."],"fun_headline_variants":["GraphSAGE on MRI DAGs tops baselines for pCR prediction","Temporal graph of DCE-MRI trajectories improves pCR over SSL","Three trajectory losses plus GNN lift pCR metrics on ISPY-2","Longitudinal ResNet graphs predict NACT response better than vision","Time-aware DAG modeling of MRI scans yields higher pCR accuracy"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That the three asymmetric losses learn clinically meaningful disease trajectories rather than artifacts of the homogeneous ISPY-2 scan schedule and the fixed ResNet18 features.","fun_headline_variants_meta":{"raw":{"variants":["GraphSAGE on MRI DAGs tops baselines for pCR prediction","Temporal graph of DCE-MRI trajectories improves pCR over SSL","Three trajectory losses plus GNN lift pCR metrics on ISPY-2","Longitudinal ResNet graphs predict NACT response better than vision","Time-aware DAG modeling of MRI scans yields higher pCR accuracy"]},"model":"grok-4.5","effort":"low","cost_usd":0.008428,"raw_usage":{"total_tokens":2012,"prompt_tokens":815,"num_sources_used":0,"completion_tokens":94,"cost_in_usd_ticks":84280000,"prompt_tokens_details":{"text_tokens":815,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1103,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":815,"tokens_out":94,"duration_ms":8301,"temperature":1.0,"reasoning_tokens":1103,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T11:32:59.036192+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain and re-evaluate the identical GNN-pCR pipeline on an external multi-center DCE-MRI cohort whose inter-scan intervals vary substantially; if balanced accuracy and AUC fall to the level of the CNN or LSTM baselines, the claim that the trajectory losses capture generalizable longitudinal dynamics is falsified.","supporting_citations":[],"review_version":1}