{"id":"50150aa3-aeef-4ffd-afef-321eb370c78f","arxiv_id":"2607.11533","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DiT-RC predicts cholangiocarcinoma PNI from tumor-centered 3D MRI with AUC 0.731 at 257.57 GFLOPs by routing attention heads, tokens, and MLP width in a diffusion classifier.","lead":"A diffusion-transformer model with adaptive routing predicts perineural invasion from 3D MRI at AUC 0.731 while cutting compute versus a non-routed variant. It targets a hard preoperative oncology signal on a small single-center cohort and needs external validation before clinical use.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection beyond the reader's correctly identified small single-center cohort limitation.","rationale":"The central claim is a methods result (DiT-RC routing yields nearly identical AUC to the full DiT-C at substantially lower FLOPs/latency while improving noise robustness over CNN/ViT baselines). All internal evidence—architecture, training objective, ablations of head/width/token routing and local 3\times3\times3 conv, budget λ_target sweep, Grad-CAM consistency—is mutually consistent and does not contain a hidden mathematical or experimental contradiction. The only premise that can still overturn the claim as stated is external validity of the 155-patient single-center private set; that premise is already identified by the reader, is explicitly acknowledged by the authors, and is reflected in the CONDITIONAL verdict and medium correctness_risk. No additional load-bearing flaw (e.g., data leakage, mis-specified FLOPs accounting, or broken diffusion scoring) is supported by the manuscript. Therefore the verdict remains CONDITIONAL with no adjustment required.","tokens_in":9290,"tokens_out":603,"duration_ms":6488,"concrete_test":"If an independent multi-center T2 MRI cohort (different scanners/protocols, n≥200 with balanced PNI labels) becomes available, retrain or fine-tune DiT-RC under identical 5-timestep two-class protocol and recompute AUC + 95% CI; if the point estimate falls outside the original 0.66–0.80 interval or the CI width exceeds 0.20, the generalizability premise fails and the clinical claim must be further restricted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest_assumption already isolates the true load-bearing premise: that 155 single-center patients (61+/94-) with tumor-centered 96\times96\times48 ROIs and patient-level 5-fold CV suffice to underwrite the reported AUC 0.731 (CI 0.66–0.80) and the efficiency claim of 257.57 GFLOPs nearly matching DiT-C. No stronger internal inconsistency appears. The diffusion scoring (Sec. II-A: my = ||ε̂\theta(xt,t,y)-ε||²₂, ŷ = arg min my), AdaLN conditioning, Gumbel–Sigmoid head/token/width routing (Eqs. 3–5), reff budget (Eq. 7), and joint loss (Eqs. 6,8) are coherent; Tables I–V and Figs. 2–3 supply consistent ablations and noise-robustness evidence. The paper itself flags multi-center validation as necessary (Conclusion). Thus the concern does not require a new attack; it is already correctly scoped.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes DiT-RC, a diffusion-transformer classifier for preoperative perineural invasion (PNI) prediction from tumor-centered 3D T2-weighted MRI. Class scores are obtained by comparing class-conditional noise reconstruction errors under a transformer denoising backbone conditioned on timestep and class hypothesis (Sec. II-A). To reduce the cost of volumetric tokens and iterative denoising, the authors introduce adaptive routing over attention heads, spatial tokens (with a lightweight 3D-convolution importance estimator), and MLP width, trained with a joint diffusion-classification and keep-rate budget loss (Eqs. 3–8). On a single-center cohort of 155 patients with patient-level 5-fold CV, DiT-RC reports AUC 0.731 (bootstrap 95% CI 0.66–0.80) at 257.57 GFLOPs, nearly matching a non-routed DiT-C (0.733 at 418.79 GFLOPs), with improved noise robustness relative to CNN and standard transformer baselines and ablations of routing components and budget targets (Tables I–V, Figs. 2–3).","tokens_in":9605,"tokens_out":1336,"duration_ms":21906,"significance":"Preoperative PNI prediction is clinically relevant for cholangiocarcinoma planning, and combining diffusion-based class scoring with long-range volumetric transformers is a reasonable response to weak, noise-sensitive peritumoral cues. The main technical contribution is the joint adaptive routing design (heads, tokens with local 3D context, MLP width) under an explicit FLOPs-proxy budget, which is well ablated and yields a clear efficiency–accuracy trade-off (Tables III–V). Strengths include patient-level stratified CV, bootstrap CI, multi-noise robustness evaluation, a U-Net diffusion baseline, and transparent reporting that multi-center validation remains necessary. If the efficiency results hold under broader validation, the work is a useful methods contribution for deployable diffusion classifiers in 3D medical imaging; absolute clinical performance claims remain limited by cohort size.","major_comments":[{"comment":"Sec. III-A and Table I: the central performance claim (AUC 0.731, best across clean/noisy settings) rests on n=155 (61+/94−) single-center patients with a bootstrap CI of 0.66–0.80. That interval substantially overlaps reported baselines (e.g., Swin 0.710, ViT 0.700, Diffusion Classifier 0.692). No paired statistical comparison (e.g., DeLong or patient-level bootstrap tests of AUC differences) is provided. Without such tests—or a clearer framing that absolute superiority is not established—the ranking narrative in Sec. III-B is stronger than the evidence supports. Please add significance testing for key pairwise AUC comparisons and temper abstract/conclusion language accordingly.","section":null},{"comment":"Sec. III-A / Table II–III: FLOPs and latency are reported “per patient prediction under the five-timestep and two-class protocol, excluding repeated noise ensembling,” while inference also averages over multiple timesteps and random seeds. For a paper whose primary selling point is efficiency, the evaluation protocol must make clear (i) which operations are included in the 257.57 GFLOPs / 140.47 ms numbers, (ii) the realized keep ratios k_attn, k_tok, k_mlp used for routed FLOPs, and (iii) how weights w_attn, w_mlp in Eq. (7) are set. Please report mean keep ratios per block/timestep and a full end-to-end cost that matches the protocol used to obtain the reported AUC.","section":null},{"comment":"Sec. II-C, Eq. (8): L_cls = L_mse^(y) + α L_CE is underspecified for a diffusion classifier. Clarify whether L_mse is computed only under the true class y, under both class hypotheses, or as a contrastive term; how reconstruction errors m_y are turned into CE targets with learnable τ; and the values (or selection) of α, λ_budget, and the Gumbel–Sigmoid temperature. These choices are free parameters of the method and affect reproducibility of both AUC and the routing budget behavior in Tables III–V.","section":null}],"minor_comments":[{"comment":"Fig. 1 caption states N=6 blocks; confirm consistency with the FLOPs model and whether depth was ablated.","section":null},{"comment":"Fig. 2 Grad-CAM comparison of DiT-RC vs DiT-C is qualitative only; a brief note on how gradients are taken through the diffusion scoring path would help.","section":null},{"comment":"Table I noise levels NL/NM/NH should explicitly map to σ∈{0.1,0.3,0.5} in the table header or footnote.","section":null},{"comment":"Related-work citations include several concurrent/self works on PNI MRI by overlapping authors; a short differentiation paragraph (what DiT-RC adds beyond LoSA-Net / MMA-Former / NeoNet) would improve novelty framing.","section":null},{"comment":"Notation: Eq. (2) writes ε̂_θ(x_t,t,y) but the surrounding text sometimes refers to reconstruction error without restating the noise target; keep a single consistent definition of m_y.","section":null},{"comment":"Minor prose: “V olumetric” spacing artifact in Sec. II-B.2; standardize “DiT-RC” vs “diffusion transformer” naming in the abstract vs body.","section":null}],"recommendation":"major_revision","confidential_remarks":"Fit is reasonable for a methods-oriented medical imaging / efficient-ViT venue. The efficiency ablations are the strongest part; the clinical performance story is under-powered. I would not reject on cohort size alone given the paper’s own multi-center caveat, but I would require statistical comparison of AUCs and a fully specified cost protocol before acceptance. Overlap with the authors’ other recent PNI-MRI submissions is worth a quick editorial check for incrementalism, not misconduct."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent methods paper that packages a diffusion transformer classifier with adaptive routing over heads, tokens, and MLP width for preoperative PNI prediction on 3D T2 MRI. It is not a conceptual leap, but the combination is cleanly executed and the efficiency numbers are honest.\n\nWhat is new is the specific DiT-RC design for this task: class-conditional noise scoring on a transformer backbone, plus joint routing of attention heads, spatial tokens (with a lightweight 3D conv for local context), and MLP width, trained under an explicit FLOPs-proxy budget. They report AUC 0.731 (bootstrap CI 0.66–0.80) at 257.57 GFLOPs, nearly matching the non-routed DiT-C (0.733 at 418.79 GFLOPs), with better noise robustness than the CNN and standard ViT/Swin baselines, and a sensible U-Net diffusion comparison. Ablations (head/token/width, local conv, λ_target sweep) and the routing maps line up with the claims. Math and training setup are coherent; citations cover DiT, DynamicViT-style sparsification, Gumbel routing, and AdaLN without obvious gaps. Writing is clear.\n\nSoft spots are real but scoped. The load-bearing limit is the private single-center cohort of 155 patients (61+/94−), tumor-centered 96×96×48 ROIs, patient-level 5-fold CV only, wide CI, and no released code/data. The authors themselves flag multi-center validation. That does not break the internal experiments; it caps how far the clinical claim can travel. Free parameters (budget target, loss weights, timestep ensemble) are standard for this style of work. Self-citations to related PNI MRI papers by the same group are present but not load-bearing for the routing result.\n\nWho it is for: people working on efficient 3D medical transformers or diffusion classifiers, and anyone doing PNI/cholangiocarcinoma imaging. A serious referee should see it; desk rejection would be the wrong call. I would bring it to reading group if we are talking efficiency or medical DiTs; I would cite the routing pattern if I am building something similar, not as settled clinical evidence.\n\nRecommendation: send to peer review. Treat external validation and data/code availability as required follow-ups, not as already settled.","headline":"Solid engineering paper: DiT + three-way adaptive routing for PNI MRI, clean ablations, small single-center n is the real limit.","tokens_in":10302,"tokens_out":596,"would_cite":true,"duration_ms":5408,"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 diffusion transformer with adaptive routing predicts perineural invasion from 3D MRI at 0.731 AUC while cutting compute.","keywords":["diffusion transformer","adaptive routing","token selection","computational efficiency","3D MRI analysis","perineural invasion","cholangiocarcinoma"],"falsifier":"An external multi-center test set of comparable size, acquired on different scanners and protocols, on which DiT-RC either drops below the CNN/transformer baselines or loses its FLOPs advantage relative to the non-routed model.","tokens_in":10166,"feed_emoji":"🧬","tokens_out":583,"duration_ms":5655,"temperature":0.7,"pith_summary":"Perineural invasion is a hard-to-see prognostic sign in cholangiocarcinoma that often lies outside the tumor itself. The paper shows that casting the prediction as class-conditional denoising inside a transformer, then dynamically routing attention heads, spatial tokens, and MLP width, can recover those subtle volumetric cues more reliably than ordinary CNNs or transformers, especially under noise. On a 155-patient single-center T2-MRI cohort the method reaches an AUC of 0.731 at 257.57 GFLOPs—nearly matching a non-routed version that costs far more. The practical claim is that the same accuracy can be obtained with roughly 40 percent fewer operations, making diffusion-transformer classifiers closer to usable in clinical settings. A sympathetic reader cares because preoperative PNI knowledge can change surgical and neoadjuvant decisions, yet current imaging models either miss the diffuse signal or are too expensive to run.","feed_headline":"Routed diffusion transformer predicts PNI at 0.731 AUC","feed_subtitle":"Adaptive head, token and MLP routing cuts FLOPs by ~40% while holding accuracy on 3D MRI","key_machinery":"DiT-RC: a diffusion transformer whose blocks use lightweight routers (conditioned on timestep and global features, plus a local 3D convolution for tokens) to binary-mask attention heads, spatial tokens, and MLP channel groups, so only informative computation is performed under a budget loss.","core_discovery":"Formulating preoperative PNI prediction as diffusion-based classification with a transformer denoising backbone, and inserting adaptive routing over attention heads, spatial tokens, and MLP width, yields an AUC of 0.731 at 257.57 GFLOPs on tumor-centered 3D T2 MRI—nearly matching the non-routed counterpart while remaining more noise-robust than CNN and standard transformer baselines.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Adaptive-routed diffusion transformer hits 0.731 AUC for PNI MRI","Head-token-MLP routing enables 0.731 AUC DiT PNI prediction","Diffusion transformer routing yields 0.731 AUC on peritumoral MRI","Adaptive routing cuts DiT FLOPs while holding 0.731 PNI AUC","Routed transformer denoising predicts cholangiocarcinoma PNI at 0.731 AUC"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That results from 155 patients at one hospital, using hand-cropped tumor regions and internal cross-validation alone, are enough to support the claimed accuracy and efficiency for real clinical use.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive-routed diffusion transformer hits 0.731 AUC for PNI MRI","Head-token-MLP routing enables 0.731 AUC DiT PNI prediction","Diffusion transformer routing yields 0.731 AUC on peritumoral MRI","Adaptive routing cuts DiT FLOPs while holding 0.731 PNI AUC","Routed transformer denoising predicts cholangiocarcinoma PNI at 0.731 AUC"]},"model":"grok-4.5","effort":"low","cost_usd":0.00543,"raw_usage":{"total_tokens":1465,"prompt_tokens":744,"num_sources_used":0,"completion_tokens":108,"cost_in_usd_ticks":54300000,"prompt_tokens_details":{"text_tokens":744,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":613,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":744,"tokens_out":108,"duration_ms":4989,"temperature":1.0,"reasoning_tokens":613,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T04:55:15.255661+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"An external multi-center test set of comparable size, acquired on different scanners and protocols, on which DiT-RC either drops below the CNN/transformer baselines or loses its FLOPs advantage relative to the non-routed model.","supporting_citations":[],"review_version":1}