{"id":"125a176c-2728-40d2-af27-31dba3886bab","arxiv_id":"2507.09158","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A U-Net variant using ReLU in the encoder and attention-based ReLU in the decoder reports 83.58% Dice versus 80.13% for baseline U-Net on thoracic vertebrae X-ray segmentation, based on one split and without released code or data.","lead":"The paper replaces the activation functions in one half of a U-Net with a learnable attention-based ReLU variant and reports a higher Dice score for segmenting thoracic vertebrae in spine X-rays. It is a small, incremental architecture tweak whose reported gain rests on a single experimental split without error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 4.1% Dice gain from decoder AReLU is confounded with a concurrent depth increase: Section 4.1 adds a down-sampling layer to the baseline, and Table 1's ablation does not isolate activation choice from depth.","rationale":"The reader's verdict is conditional on run-to-run variance. I agree that the lack of repeated runs and seeds is a serious weakness and that the p-value claim is not backed by distributions. My stress-test pass adds a more direct confound: the manuscript changes two architectural properties at once. Section 4.1 explicitly adds a down-sampling layer as part of the novel model, while the baseline description in Section 3.3 gives no layer count. Table 1's ablation is ambiguous about which model is row 0. If row 0 is the shallow standard U-Net, the 80.13-to-83.58 comparison includes depth and activation simultaneously. If row 0 is the deep all-ReLU model, the paper never reports the depth effect separately. In either reading, the strongest claim that swapping activation functions by path yields the improvement is not established by the present experiments. The paired t-test only shows the final models differ on this split, not why. This motivates a 2x2 factorial ablation with multiple seeds; it would settle whether activation choice or depth/capacity explains the gap. Since the data and code are not required by the paper, the conditional verdict should remain: accept only if such an ablation and repeated-run statistics are supplied. My concern is complementary to the reader's rather than identical, hence partial agreement.","tokens_in":9711,"tokens_out":9790,"duration_ms":112974,"concrete_test":"Run a 2x2 factorial ablation on the BUU AP-thoracic data: (1) baseline U-Net (with its actual number of down/up layers stated) using ReLU everywhere; (2) baseline depth with AReLU in the decoder; (3) 6-down/5-up all-ReLU; (4) the full proposed 6-down/5-up with AReLU in the decoder. Use at least five random seeds per cell, keeping the same 240/48/60 split and augmentation pipeline. Report mean plus/minus standard deviation Dice on the held-out 60-image test set. If model (3) is within noise of model (1) and model (4) is about 3.5 points above model (3), the activation story survives. If model (3) is about 3.5 points above model (1), the improvement is attributable to depth/capacity, not to AReLU. Also report whether the baseline already uses six encoder blocks and five decoder blocks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that applying AReLU in the decoder is what buys the 3.45-point Dice gain (80.13 to 83.58, Tables 1 and 3). That attribution is not isolated empirically. Section 4.1 says the novel model has six down-sampling and five up-sampling layers and explicitly motivates the inclusion of an additional down-sampling layer as part of the proposed architecture. The baseline U-Net in Section 3.3 is described only as a standard U-Net with ReLU; its layer count is never given. The layer-by-layer AReLU progression in Table 1 therefore cannot tell the reader whether row 0 is the shallow baseline with ReLU everywhere or a deeper all-ReLU variant. If row 0 is the shallow baseline, the comparison varies both depth and activation. If row 0 is the deep all-ReLU model, the contribution of the extra down-sampling layer relative to the standard U-Net is unreported, and the Table 3 comparison is not a pure activation comparison. In both readings, the conclusion that integration of dual activation functions resulted in a 4.1% improvement over standard U-Net overstates what the experiments show. The p-values in Table 4 compare final models only; they do not separate the two factors. A second, independent limitation is that all numbers come from one random split and one training run (Sections 3.2, 5.2, 5.3), so the size and stability of the effect are also unquantified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modified U-Net, called a 'Sandwich U-Net', that uses ReLU activations in the encoder/contracting path and AReLU activations in the decoder/expanding path for segmenting thoracic vertebrae in AP-view X-ray images. On a 60-image held-out test set, the authors report a Dice score of 83.58% for the proposed model versus 80.13% for a standard U-Net, along with a layer-by-layer ablation of incremental AReLU adoption and paired t-test p-values below 0.001. The paper concludes that integrating dual activation functions improves segmentation accuracy and contour quality for vertebral contouring.","tokens_in":9982,"tokens_out":5382,"duration_ms":55335,"significance":"If the reported improvement is genuine and attributable to the activation change, the paper demonstrates a simple and low-cost architectural modification that yields a modest but consistent improvement of roughly 3.5 Dice points on a clinically relevant segmentation task. The use of a public dataset and a held-out test set is a positive feature, and the layer-by-layer ablation in Table 1 is a useful presentation of the effect of progressively introducing AReLU. However, the significance is currently limited by the absence of variance estimates, the unresolved confound between activation choice and network depth, and the incomplete statistical reporting; the contribution would be substantially strengthened by an isolated ablation and rigorous multi-run evaluation.","major_comments":[{"comment":"The central claim that AReLU in the decoder produces the 3.45-point Dice improvement is confounded with a depth change: Section 4.1 states that the proposed architecture includes 'six down-sampling layers and five up-sampling layers' and explicitly justifies 'the inclusion of an additional down-sampling layer' as part of the proposal, while Section 3.3 does not specify the number of down-sampling layers in the standard U-Net. Consequently, Table 3 compares models that differ in both depth and activation, and Table 1's row for '0 AReLU' is ambiguous as to whether it represents the shallow baseline or a deeper all-ReLU variant. Please provide an ablation that fixes network depth while varying only the activation functions, and report the layer counts of all compared models.","section":"Sections 4.1, 5.4, Table 1 and Table 3"},{"comment":"The paired t-test results are not backed by a described statistical procedure. Table 3 provides a single aggregated Dice score per model; if the test is performed on per-image Dice scores, the paper must describe the pairing, report the number of paired samples, and give the mean and standard deviation (or confidence intervals) of the per-image scores for each model. If only one score per model exists, a paired t-test cannot be applied and the reported p-values are unsupported. Please either supply the full statistical details or remove the significance claims.","section":"Section 5.5, Table 4"},{"comment":"All results are based on a single random 80/20 train/validation split and one training run per configuration, with no seeds, repeated runs, or variance estimates reported. Since training stochasticity and split choice can shift Dice scores by several points, the reported 3.45-point gap cannot be distinguished from run-to-run variation. Please report the mean and standard deviation over multiple runs (e.g., five random seeds) for at least the baseline and proposed models, or provide bootstrap confidence intervals.","section":"Sections 3.2, 5.2, 5.3"},{"comment":"The grid search for alpha and beta initializations is reported as selecting the values that give 'the best accuracy', but the manuscript does not state whether this selection was made on the validation set or the test set. If the test set was used for selection, the reported test Dice of 83.58% is optimistically biased by selection. Please clarify that the hyperparameter selection was performed solely on validation data, and if not, re-evaluate on a truly held-out test set.","section":"Section 4.4, Table 2"}],"minor_comments":[{"comment":"The sentence 'the second half of the \"U\" shown in Fig. 3 depicts the novel U-Net architecture' is grammatically incomplete; it should be split into two sentences for clarity.","section":"Section 4.1"},{"comment":"Describing AReLU as 'attention based' may be misleading because the equations define a learnable parametric activation function, not an attention mechanism over feature maps; please clarify the relationship to reference [5] and justify the terminology.","section":"Sections 1 and 4.3"},{"comment":"The text says 'C(·) clamps the input variable into [0.01, 0.99]', but in the equation C is applied to the parameter alpha rather than to the input xi; please reword to avoid confusion.","section":"Equation (1)"},{"comment":"The text 'Python III' should be 'Python 3', and 'google co-lab' should be 'Google Colab' for correctness.","section":"Section 5.2"},{"comment":"The manuscript gives no training hyperparameters (learning rate, batch size, number of filters, etc.) for any of the models, and it does not include a code or data availability statement; such details are needed for reproducibility.","section":"Sections 5.2 and 5.3"},{"comment":"Tables 1 and 2 report Dice scores without standard deviations or an indication of whether the numbers are from the validation or test set; adding these details would help readers judge the reliability of the comparisons.","section":"Tables 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is plausible and the public dataset allows the authors to address the main concerns experimentally. However, the confound between activation change and depth change, the lack of variance estimates, and the unsupported paired t-test are load-bearing issues that must be fixed before publication. I do not see a reason for rejection if the authors can provide the missing evidence, but the current manuscript cannot support the claimed 4.1% improvement as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper reports a modest, plausible empirical gain—roughly 4–4.1 Dice points—by using AReLU in the decoder of a U-Net for thoracic vertebra segmentation on AP X-ray. The method is an incremental combination of known pieces: U-Net plus AReLU plus one extra down-sampling block. That is not a new class of method, but if the numbers hold, it is a useful data point for a specific clinical contouring task.\n\nWhat it does well: the dataset is public, the task is concrete and relevant, and the authors are explicit in Section 7 that they did not compare with recent SOTA methods. The layer-by-layer ablation in Table 1 is suggestive—Dice rises monotonically as decoder layers switch to AReLU—and the figures give a qualitative sense that edge contours improve.\n\nWhere it falls short, in order of importance:\n\n1. The confound flagged in the stress-test note is real. Section 4.1 introduces an additional down-sampling layer as part of the proposed architecture, while Section 3.3 never specifies the baseline U-Net's depth. The headline comparison, 80.13 to 83.58, therefore varies both depth and activation. Table 1's row 0 could be the shallow baseline or a deeper all-ReLU variant; the text does not separate the two factors. As reported, the paper overstates what the experiments show.\n\n2. The statistical claim is unsupported. Table 4 reports paired t-test p-values below 0.001, but there is no description of what was paired, no per-image scores, no repeated runs, and no seeds. With one Dice number per model, the test as described is not reproducible.\n\n3. Single split, single run. All numbers come from one random 80/20 train-validation split and one training run per configuration (Sections 3.2, 5.2, 5.3). Even setting aside the depth confound, the size and stability of the effect are unquantified.\n\n4. The alpha/beta selection is mild and acceptable: it was done on validation before touching the test set. That part is not a red flag.\n\nAll of this is fixable. The authors need to rerun with a fixed depth, report per-image Dice across multiple seeds and splits, and describe the paired test properly. The paper is clearly written and the limitation section is honest, so I would not desk-reject it; a serious referee could quickly tell the authors exactly what to add. I would not cite it in its current form, and I would not bring it to a reading group until the ablation confound is resolved.","headline":"A plausible but not-yet-isolated empirical claim: the 4.1% Dice gain attributed to the ReLU/AReLU sandwich is confounded with an added down-sampling layer, and the statistics as reported cannot be checked.","tokens_in":10601,"tokens_out":2390,"would_cite":false,"duration_ms":30008,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a U-Net with ReLU down and AReLU up reaches 83.58% Dice on thoracic X-ray vertebra contouring, beating 80.13% for a standard U-Net.","keywords":["U-Net","AReLU","ReLU","vertebra segmentation","thoracic spine X-ray","Dice score","semantic segmentation","activation function"],"falsifier":"Retrain the baseline U-Net and the sandwich U-Net on the same 300 anterior-posterior X-rays across several random splits and several seeds per split, then compare Dice distributions on one fixed 60-image test set; if the two distributions overlap substantially or the mean difference is not consistently above zero with non-overlapping confidence intervals, the claimed superiority fails.","tokens_in":9499,"feed_emoji":"🩻","tokens_out":9469,"duration_ms":102455,"temperature":0.7,"pith_summary":"Outlining individual vertebrae in spinal X-rays is done by hand today, which is slow and inconsistent, and automated contouring would help mobility assessment and surgical planning. This paper tries to establish that a small change to the standard U-Net—plain ReLU in the contracting half and attention-based ReLU (AReLU) in the expanding half—makes thoracic vertebra contouring more accurate. On 60 held-out anterior-posterior X-ray images from a public spine dataset, the proposed 'sandwich' U-Net reports a Dice score (the standard 0–100 overlap measure) of 83.58% versus 80.13% for a standard U-Net, with paired t-test p-values below 0.001 against all three comparators. A layer-by-layer experiment shows the score climbing as each decoder block is switched to AReLU, which the authors attribute to better preservation of vertebral edges and partial vertebrae. If the result holds, the improvement costs almost nothing in architecture complexity and could be folded into existing segmentation pipelines.","feed_headline":"Spine X-ray vertebra contour Dice jumps from 80.1% to 83.6%","feed_subtitle":"On a held-out 60-image set, a U-Net with ReLU down and AReLU up beats the plain U-Net and two other baselines.","key_machinery":"The machinery is the sandwich U-Net itself: an encoder–decoder network with six ReLU-activated down-sampling blocks and five AReLU-activated up-sampling blocks. AReLU is the load-bearing component, a learnable activation that keeps negative inputs alive with a small clamped slope while scaling positive inputs with a sigmoid-driven gain, giving the decoder a trainable way to emphasize or suppress features during reconstruction. The paper's incremental experiment—adding AReLU to one additional decoder block at a time and watching the Dice score rise monotonically—is the evidence that ties this mechanism to the contour improvements.","core_discovery":"At its core, the paper claims that separating activation functions along the two halves of a U-Net makes each half specialize: the six down-sampling blocks, using ReLU, extract sparse, stable features, while the five up-sampling blocks, using AReLU, adaptively reconstruct details and sharpen boundaries. AReLU is a learnable attention-based ReLU defined as $F(x_i,\\alpha,\\beta)=R(x_i)+L(x_i,\\alpha,\\beta)$, where the negative branch has a clamped slope $C(\\alpha)\\in[0.01,0.99]$ and the positive branch is scaled by $1+\\sigma(\\beta)$. With $\\alpha=\\beta=0.9$, the model achieves 83.58% Dice on the 60-image test set versus 80.13% for the baseline U-Net, and the paper's step-by-step table shows Dice rising from 80.13% to 83.58% as zero through five decoder blocks are converted to AReLU. The authors conclude that the sandwich model produces contours that match ground truth more closely, especially at lower vertebral borders, reducing over- and under-segmentation.","pith_inferences":["The paper changes two things at once—the decoder activations and the number of down-sampling blocks—so the AReLU contribution alone is not isolated; a follow-up that deepens the baseline equally would separate the two effects.","A transfer test on lateral spine X-ray views or CT slices would show whether path-split activation is a general encoder–decoder property or specific to anterior-posterior thoracic images.","Claimed edge benefits would be better captured by boundary metrics like Hausdorff distance or average surface distance, since Dice is fairly insensitive to small contour shifts.","Replacing AReLU with a simpler learnable slope such as PReLU in the decoder would reveal whether the attention gating or merely the learnable slope drives the gain."],"forward_implications":["A team with an existing U-Net could adopt the change by swapping only the decoder activation functions, with no new data collection and no added inference-time memory cost beyond the AReLU computation.","At the reported accuracy, automated contours could support quantitative measurements of vertebral alignment, curvature, and spacing for early diagnosis of scoliosis and kyphosis and for planning spinal fusion.","The reported p-values below 0.001 for SegNet, EfficientNet, and U-Net comparisons are the paper's evidence that the Dice gap is not a within-test chance fluctuation.","Because the improvement appears most strongly at vertebral borders and partial vertebrae, the method targets exactly the cases where manual contouring is slowest and most inconsistent."],"supporting_citations":[{"why":"Defines the AReLU activation function used throughout the up-sampling path.","marker":"[5]"},{"why":"Supplies the evidence that AReLU enables faster training and resists gradient vanishing.","marker":"[4]"},{"why":"SegNet is one of the three baselines whose Dice score the proposed model must beat.","marker":"[1]"},{"why":"EfficientNet is another baseline architecture in the comparison table.","marker":"[32]"},{"why":"Establishes U-Net as the standard medical-segmentation architecture that the work modifies.","marker":"[6]"}],"fun_headline_variants":["Sandwich U-Net lifts spine X-ray contour Dice by 4.1%","Dual-activation U-Net sharpens vertebra contours on X-ray","ReLU down, AReLU up: sandwich net improves spine contour Dice","Spine X-ray segmentation: sandwich U-Net hits 83.6% Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one random 80/20 split of the 300-image set and one training run per configuration are enough to trust the 3.5-point Dice gap as a real difference rather than split choice or run-to-run noise.","fun_headline_variants_meta":{"raw":{"variants":["Sandwich U-Net lifts spine X-ray contour Dice by 4.1%","Dual-activation U-Net sharpens vertebra contours on X-ray","ReLU down, AReLU up: sandwich net improves spine contour Dice","Spine X-ray segmentation: sandwich U-Net hits 83.6% Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000339,"raw_usage":{"total_tokens":1884,"prompt_tokens":972,"completion_tokens":912,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":826}},"tokens_in":588,"tokens_out":912,"duration_ms":9044,"temperature":1.0,"reasoning_tokens":826,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:01:24.732813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the baseline U-Net and the sandwich U-Net on the same 300 anterior-posterior X-rays across several random splits and several seeds per split, then compare Dice distributions on one fixed 60-image test set; if the two distributions overlap substantially or the mean difference is not consistently above zero with non-overlapping confidence intervals, the claimed superiority fails.","supporting_citations":[{"cited_title":"AReLU: Attention-based Rectified Linear Unit","cited_arxiv_id":"2006.13858","evidence_quote":"Defines the AReLU activation function used throughout the up-sampling path."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the evidence that AReLU enables faster training and resists gradient vanishing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SegNet is one of the three baselines whose Dice score the proposed model must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EfficientNet is another baseline architecture in the comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes U-Net as the standard medical-segmentation architecture that the work modifies."}],"review_version":1}