{"id":"6aac1f79-02d6-479d-9973-d8f7e4830308","arxiv_id":"2505.24421","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A multi-encoder, augmentation-aware architecture with attention-weighted fusion (MEAL-BD) is claimed to improve robustness of 3D CT-to-T1 MRI translation over single-stream baselines.","lead":"pyMEAL is a deep learning toolbox that gives each type of image augmentation its own encoder pathway and then fuses the resulting features before decoding, for CT-to-MRI translation. The authors report that the controller-weighted variant, MEAL-BD, beats their other model variants on PSNR and SSIM, though the comparison is confounded by architecture size and missing external baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BD's 273M parameters versus 4-14M baselines confound the reported gains; the 75% parameter-reduction claim in §2.1.4 is contradicted by §2.5, so the central robustness advantage is not established.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the evaluation is not fair because BD has vastly more parameters and possibly receives different test inputs than the baselines. The paper itself provides the evidence for this concern in Section 2.5, where BD's parameter count is 273,265,538 versus 4,428,545 for NA/TA and 13,725,953 for CC. This directly contradicts the 'reduces parameters by 75%' statement in Section 2.1.4, which is not a minor typo: it invalidates the claim that the performance advantage comes from the augmentation-aware fusion mechanism. Without a parameter-matched control or an explicit ablation holding capacity fixed, higher PSNR and SSIM cannot be attributed to MEAL-BD's architecture. The paper also does not compare against established CT-to-MRI translation methods, and the robustness evaluation uses the same synthetic augmentations used in training, so the generalizability claim rests on a narrow protocol. The underlying idea is worth testing, but the reported evidence is insufficient for acceptance. Therefore the reader's REJECT verdict remains appropriate; I do not see a reason to move the verdict in either direction.","tokens_in":15806,"tokens_out":3195,"duration_ms":39223,"concrete_test":"Reproduce Section 2.5's parameter counts from the released GitHub/PyPI code, then run an equalized-capacity ablation: (a) reduce BD to roughly 4.4M parameters by enforcing the shared encoder described in Eq. 4; (b) scale NA/TA/CC/FL to roughly 273M parameters by increasing channel widths or repeating residual blocks; use identical train/test splits, augmentation protocol, and optimizer settings. If parameter-matched BD does not beat all baselines in PSNR/SSIM on the same test sets, the central robustness claim is not supported. Also report BD's test-time inputs (one raw volume versus four augmented volumes) to verify information equivalence across methods.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is capacity confounding. The central claim is that MEAL-BD's multi-encoder augmentation-aware fusion, not raw model size, causes consistent PSNR/SSIM gains. Section 2.1.4 states that BD uses a single shared encoder f_theta and 'reduces parameters by 75% compared to other methods'. Section 2.5 reports BD: 273,265,538 trainable parameters; CC: 13,725,953; FL: 10,760,577; NA/TA: 4,428,545. That is between 6x and 62x larger, not 75% smaller. A model with 60x greater capacity would be expected to fit the training distribution better and achieve higher fidelity regardless of the controller or fusion design. The paper provides no parameter-matched baselines and no ablation isolating the controller's contribution. In addition, the test-time input protocol for single-stream baselines is unspecified, leaving open the alternative explanation that BD receives four augmented inputs and a much larger hypothesis space. If the parameter table is accurate, every reported 'BD significantly outperforms' result is confounded; if it is a typo, it still means the submitted evidence does not support the claimed parameter efficiency or the architectural attribution of the gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes pyMEAL, a multi-encoder augmentation-aware learning framework for 3D medical image translation, evaluated on CT-to-T1-weighted MRI synthesis using OASIS-3 data. The framework routes four augmentation variants through parallel encoder streams and compares three fusion strategies: concatenation (CC), a fusion layer (FL), and a dynamic controller block (BD). The central claim is that MEAL-BD, by dynamically weighting augmentation-specific features, consistently outperforms the other configurations and baseline no-augmentation/traditional-augmentation models in PSNR and SSIM on both predefined and unseen test data.","tokens_in":15961,"tokens_out":4921,"duration_ms":58974,"significance":"The core idea—treating augmentations as complementary views and learning to fuse them with an attention-style controller—is timely and potentially useful for robustness in medical image translation. The paper is also commendable for releasing pip-installable code and for applying a detailed statistical protocol (normality tests, Kruskal-Wallis, Dunn with Bonferroni). However, the empirical evidence as presented is confounded. The reported trainable-parameter counts contradict the claimed parameter reduction, the evaluation lacks external state-of-the-art baselines, and the test-time input protocol for single-stream baselines is unspecified. If these issues were resolved with controlled, parameter-matched comparisons and full pairwise reporting, the contribution could be solid; in its current form, the central claim is not established.","major_comments":[{"comment":"The claim in Section 2.1.4 that 'This architecture reduces parameters by 75% compared to other methods' is directly contradicted by the parameter counts in Section 2.5: BD has 273,265,538 trainable parameters, whereas CC has 13,725,953, FL has 10,760,577, and NA/TA have 4,428,545. BD is therefore 6 to 62 times larger than the alternatives, not 75% smaller. Because model capacity is a well-known confound for reconstruction quality, the reported PSNR/SSIM advantages of BD cannot be attributed to the controller or fusion mechanism without parameter-matched baselines or ablations that isolate the architectural contribution.","section":"Section 2.1.4 and Section 2.5"},{"comment":"The claim that MEAL-BD 'consistently outperformed competing approaches' is not supported by the reported pairwise tests. Under the intensity augmentation condition in Table 2, Dunn's tests are reported for BD vs FL and BD vs TA, but not for BD vs CC or BD vs NA, even though the mean scores show BD, NA, and CC close together (21.210, 20.590, 20.054). Similarly, in the no-augmentation results in Table 1, only BD vs TA is reported, not BD vs CC, FL, or NA. The central claim of consistent superiority requires these key pairwise comparisons to be reported; their absence raises the possibility that the differences are not significant.","section":"Section 3.2.1 and Table 2"},{"comment":"All comparative results are obtained against the paper's own internal configurations (NA, TA, CC, FL). The Introduction states that 'MEAL outperforms state-of-the-art single-stream and multi-modal baselines,' but no existing CT-to-MRI translation method (e.g., CycleGAN, Pix2Pix, or published U-Net variants) is evaluated. Without at least one external baseline, the robustness and superiority claims are not contextualized against the literature, and the reader cannot judge whether the gains are specific to this dataset and setup or representative of genuine improvement.","section":"Section 3 evaluation and Section 1"},{"comment":"The test-time evaluation protocol under augmentation-specific conditions is not specified. In Section 3.2.1, results are presented 'under various augmentation scenarios (rotation, crop, flip, and intensity),' but it is not stated whether all models receive the same augmented input at test time, or whether the multi-encoder models (BD, CC, FL) receive four augmented input views while the single-stream models (NA, TA) receive only one. If the input information differs across methods, the comparison is not fair, and the claimed advantage of BD could partly stem from receiving multiple views rather than from the fusion architecture. The exact input protocol for each method at test time must be described.","section":"Section 3.2 and Section 2.2"}],"minor_comments":[{"comment":"The text states 'L1 the mean square error,' but L1 conventionally denotes mean absolute error. Either the notation or the definition should be corrected to avoid ambiguity.","section":"Section 2.3, Eq. (12)"},{"comment":"Equation (2) uses the symbol M for channel-wise concatenation, while the text and symbols list define L as the concatenation operator. Please use a single consistent notation.","section":"Section 2.1.2, Eq. (2)"},{"comment":"There are minor typos, including 'hightligts' in Section 3.5 and 'The, we define' in Section 2.3. A careful proofreading pass is needed.","section":"Section 3.5 and Section 2.3"},{"comment":"The caption refers to 'Multi-Stream with a Build Controller method,' while the text and acronym list define BD as 'Builder Block.' The terminology should be made consistent.","section":"Figure 2"},{"comment":"Several p-values are reported as 0.0000. These should be reported as p < 0.0001 or as exact values with a clear convention.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a serious internal contradiction: the claimed 75% parameter reduction in Section 2.1.4 is incompatible with the parameter table in Section 2.5. The authors must either correct the parameter counts or provide parameter-matched baselines. The missing pairwise comparisons for key claims and the absence of external baselines are also significant issues. The paper is not ready for publication in its current form, but the core idea is worthwhile and the code availability is a plus; with careful additional experiments and clarifications, a resubmission could be viable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a sensible architectural idea—multi-encoder augmentation-aware fusion with a learned controller—but its central evidence is undermined by a parameter-count contradiction. BD has 273M parameters versus 4-14M for the other variants, yet Section 2.1.4 claims a 75% parameter reduction. That confounds the headline gains, and the paper also lacks external baselines and an explicit test-time input protocol.\n\nWhat's genuinely new: applying augmentation-aware multi-stream encoding with three fusion strategies (concatenation, fusion layer, dynamic controller) to 3D CT-to-T1 translation. That's a legitimate extension of AugSelf/CASSLE-style ideas, and the authors run a proper within-paper comparison with standard statistical tests. They also ship code and models, which is more than many papers do.\n\nThe soft spots are real and load-bearing. The parameter mismatch is not a minor copy-paste issue: if the table is accurate, BD has 6-60x the capacity of its comparators, so the reported PSNR/SSIM gains could be entirely from model size. If the table is wrong, then the evidence for the architecture's benefit is missing because the reported numbers aren't attributable. Either way, the claims as written don't hold. The test-time protocol is also unspecified: it's unclear whether BD sees four augmented views at inference while NA/TA see one, which would be an information advantage. And 'robustness' is tested with the same geometric/intensity perturbations used in training, on the same OASIS-3 dataset, so the generalization claim is overstated. No comparison to established CT-to-MRI translation methods.\n\nOn the plus side, the paper is honest about its own limitations in the discussion (parameter growth, need for head pruning), and the code release means a competent reader could re-run with parameter-matched baselines. The idea deserves further testing.\n\nWho this is for: researchers working on robust medical image synthesis who want a concrete architecture to build on. It deserves a serious referee, but the current submission is not acceptable as-is. I'd send it to peer review with a clear request for major revision: fix the parameter accounting, add parameter-matched baselines, compare to at least one external method, and specify the test-time input protocol. If the authors can't resolve the capacity confound, reject.","headline":"A sensible multi-encoder augmentation-aware architecture for CT-to-T1, but the headline gains are confounded by a 273M-vs-4-14M parameter disparity and no external baselines.","tokens_in":16604,"tokens_out":3536,"would_cite":false,"duration_ms":40099,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68U10","92C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that treating each augmentation as a distinct anatomical view, processed by its own encoder path and reweighted by a learned controller, makes CT-to-T1 MRI translation robust to geometric and intensity perturbations.","keywords":["medical image translation","data augmentation","multi-encoder architecture","attention weighting","CT-to-MRI","3D image reconstruction","PSNR","SSIM"],"falsifier":"Run a controlled comparison in which every model receives the same four augmented views (for instance, as a four-channel input or through a shared stem) and the total trainable parameter count is matched to MEAL-BD's, then evaluate PSNR and SSIM on the same unseen test set under rotation, crop, flip, and intensity perturbations. If a single-stream baseline matches or beats MEAL-BD, the claim that controller-weighted fusion is responsible for the robustness fails.","tokens_in":15536,"feed_emoji":"🧠","tokens_out":5707,"duration_ms":68332,"temperature":0.7,"pith_summary":"The paper tries to establish that data augmentations should be treated as complementary anatomical views rather than as noisy perturbations, and that a multi-encoder architecture with a learned controller can exploit them for robust 3D medical image translation. It proposes MEAL, which runs flip, rotation, crop, and intensity variants through dedicated encoder pathways and fuses the resulting features before decoding. In CT-to-T1-weighted MRI translation, the controller-weighted variant MEAL-BD reports the highest PSNR and SSIM on both predefined and unseen test sets, under geometric perturbations and under standard imaging conditions. If the claim holds, augmentation-aware feature fusion is a design principle that could improve robustness in other medical imaging tasks where acquisition conditions vary.","feed_headline":"Four augmented views beat one in CT-to-MRI translation","feed_subtitle":"A controller that reweights augmentation-specific features keeps PSNR and SSIM high under rotation, crop, flip, and intensity shifts.","key_machinery":"The load-bearing component is the controller block that computes per-stream attention weights $\\alpha_k = \\mathrm{softmax}(w^\\top \\mathrm{ReLU}(W \\cdot \\mathrm{GAP}(h_k)))$ and fuses shared-encoder features as $F_{\\mathrm{BD}} = \\sum_{k=1}^4 \\alpha_k f_\\theta(A_k(X))$, where $A_k$ are differentiable flip, rotation, crop, and intensity modules embedded in the computational graph. The shared encoder $f_\\theta$ processes all four augmented views, and the decoder $\\Gamma_\\phi$ reconstructs the translated volume from the weighted combination. This mechanism lets the network suppress uninformative transformations and amplify informative ones during training, which the paper identifies as the source of robustness.","core_discovery":"MEAL-BD, the variant with an adaptive controller block, consistently outperforms the other configurations (no augmentation, traditional augmentation, concatenation, and fusion-layer fusion) in CT-to-T1-weighted MRI translation. On unseen test data it reports the highest mean PSNR of 23.03 dB and SSIM of 0.733, and on the predefined test set 24.12 dB and 0.745; under rotation, crop, and flip augmentations it stays above roughly 23 dB and 0.72, while intensity perturbation yields lower but still leading scores. The authors attribute the gain to dynamic attention weighting of augmentation-specific features rather than to simple concatenation or averaging of those features.","pith_inferences":["The paper does not specify whether single-stream baselines received one or four augmented views at test time; if they received one while BD received four, part of the gain could be input diversity rather than the controller.","The parameter counts in Section 2.5 (BD: 273M; CC/FL: 10–14M; NA/TA: 4.4M) contradict Section 2.1.4's claim of a 75% parameter reduction, so the efficiency advantage is not established.","A natural testable extension is to replace the fixed set of four augmentations with a learned augmentation policy, where the differentiable modules $A_k$ are optimized by the same gradient signal.","The paper mentions Dice on white and gray matter segmentation as an evaluation in Section 2.4 but reports no Dice results; reporting those numbers would test whether PSNR and SSIM gains translate to anatomical accuracy."],"forward_implications":["If MEAL-BD's advantage is real, augmentation-aware controller weighting becomes a design option for other 3D medical translation tasks, not just CT-to-MRI.","The reported stability under rotation, crop, and flip suggests the architecture tolerates patient positioning and field-of-view differences without retraining.","MEAL-BD's lead on unseen data implies the learned weighting transfers to scans the model never saw during training.","The framing of augmentation as complementary views could be carried to tasks like segmentation and registration, as the paper itself proposes."],"supporting_citations":[{"why":"Supplies the paired CT and T1-weighted MRI volumes used for training, validation, and testing.","marker":"[32]"},{"why":"Provides the augmentation-aware representation learning idea that motivates treating augmentations as distinct views rather than noise.","marker":"[23]"},{"why":"The deep-learning framework used to implement and train all model variants.","marker":"[35]"},{"why":"The registration pipeline that aligns CT and MRI pairs before translation.","marker":"[36, 37]"},{"why":"Provides the 3D PSNR and SSIM metrics used for evaluation.","marker":"[38]"},{"why":"Represents the contrastive view of augmentation as data inflation, which the paper argues against.","marker":"[26]"}],"fun_headline_variants":["Adaptive controller improves CT-to-MRI translation robustness","Augmentation-aware weighting enhances medical image translation","Multiple augmented views with adaptive fusion beat single pipeline","Adaptive feature weighting outperforms in CT-to-MRI translation","Controller-weighted augmented views improve MRI synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison is fair: all methods get informationally equivalent inputs, and MEAL-BD's edge comes from the controller fusion rather than from receiving four augmented views while single-stream baselines receive one, or from having roughly 20 to 60 times more trainable parameters.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive controller improves CT-to-MRI translation robustness","Augmentation-aware weighting enhances medical image translation","Multiple augmented views with adaptive fusion beat single pipeline","Adaptive feature weighting outperforms in CT-to-MRI translation","Controller-weighted augmented views improve MRI synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001375,"raw_usage":{"total_tokens":5577,"prompt_tokens":955,"completion_tokens":4622,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":4550}},"tokens_in":571,"tokens_out":4622,"duration_ms":49358,"temperature":1.0,"reasoning_tokens":4550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:22:28.271128+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison in which every model receives the same four augmented views (for instance, as a four-channel input or through a shared stem) and the total trainable parameter count is matched to MEAL-BD's, then evaluate PSNR and SSIM on the same unseen test set under rotation, crop, flip, and intensity perturbations. If a single-stream baseline matches or beats MEAL-BD, the claim that controller-weighted fusion is responsible for the robustness fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the paired CT and T1-weighted MRI volumes used for training, validation, and testing."},{"cited_title":"Przewięźlikowski, M","cited_arxiv_id":null,"evidence_quote":"Provides the augmentation-aware representation learning idea that motivates treating augmentations as distinct views rather than noise."},{"cited_title":"Abadi, P","cited_arxiv_id":null,"evidence_quote":"The deep-learning framework used to implement and train all model variants."},{"cited_title":"Montalt-Tordera, J","cited_arxiv_id":null,"evidence_quote":"Provides the 3D PSNR and SSIM metrics used for evaluation."},{"cited_title":"Sandfort, K","cited_arxiv_id":null,"evidence_quote":"Represents the contrastive view of augmentation as data inflation, which the paper argues against."}],"review_version":1}