{"id":"1b6b99af-e496-4ad4-b3e8-12a6f18de181","arxiv_id":"2506.00434","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using 50 training epochs and one cross-validation fold, ACS-based nnU-Net with ImageNet-pretrained 2D-to-3D transfer reaches BraTS Dice scores within 0.0003 of a five-fold nnU-Net ensemble while using about 21% of the parameters.","lead":"This paper adapts the nnU-Net brain tumor segmentation model to use axial, coronal, and sagittal 2D convolutions initialized from ImageNet weights, cutting training time and parameters while keeping accuracy close to the original. It reports that these lightweight models match or slightly beat a standard five-fold ensemble on BraTS benchmarks, which matters because full nnU-Net training is expensive.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'outperform ensemble' claim rests on a quoted CV [12] baseline from the BraTS 2017 challenge, not BraTS 2018; on the valid BraTS2020 comparison the method is slightly below the ensemble (0.85546 vs 0.85580), and the 0.00065 BraTS2018 advantage is smaller than the reported run-to-run SD.","rationale":"The reader's weakest assumption already identified the quoted published baselines as not directly comparable to the single-fold, 50-epoch runs. My stress-test sharpens that concern into a potentially decisive citation/dataset mismatch in Section 5.1: reference [12] is a BraTS 2017 challenge paper, yet the authors label the row as a BraTS2018 validation result. If the quoted number is from BraTS2017, the central BraTS2018 comparison is void, and the only valid external ensemble comparison is BraTS2020, where ACS-ResNet18+BN+DA+Post achieves 0.85546 versus CV [11] at 0.85580, i.e., slightly below the ensemble. The internal comparison against Baseline+BN+DA shows only a 0.00097 Dice gain with a reported standard deviation of 0.00194, so the effect is not shown to be statistically significant. This does not invalidate the entire paper: the ACS architecture, the parameter reduction to roughly 21% of baseline, and the 50-epoch training schedule are still potentially useful contributions. But the headline claim of outperforming the ensemble is not currently supported by the evidence as written. Because the reader already issued a CONDITIONAL verdict and this concern does not move it, I recommend UNCHANGED, with the concrete test providing a way to settle the question.","tokens_in":10267,"tokens_out":7104,"duration_ms":67251,"concrete_test":"Locate the exact source table for the 'CV [12]' row (Dice 0.85600, HD95 5.25667) in Isensee et al. [12] and verify whether it reports BraTS2018 or BraTS2017 validation. Then, on the same first-fold split used by the authors, rerun the baseline nnU-Net and ACS-ResNet18-All+BN+DA with 50 epochs, identical batch normalization, data augmentation, and post-processing, evaluating on the same official BraTS2018 validation set with at least three seeds. If the ACS advantage over the rerun baseline is below one standard deviation, or the quoted baseline is not from BraTS2018, the 'outperform ensemble' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Abstract and Section 5.1 is that fast-training ACS-ResNet18 variants 'comparable or even outperform the ensemble of cross-validation models.' On BraTS2018, ACS-ResNet18-All+BN+DA obtains Dice 0.85665 and HD95 4.87734, exceeding the row labeled 'CV [12]' (0.85600 Dice, 5.25667 HD95). But reference [12] is the BraTS 2017 challenge paper ('Contribution to the BraTS 2017 Challenge'), while the text describes this as a BraTS2018 validation result. If the quoted numbers were produced on BraTS2017 validation, or under a different post-processing protocol, the BraTS2018 ensemble comparison is invalid. On BraTS2020, the same method with post-processing reaches Dice 0.85546 versus CV [11] at 0.85580, so it does not outperform the ensemble on that dataset. The residual BraTS2018 advantage over the quoted ensemble is 0.00065 Dice, below the reported standard deviation of 0.00194, and no significance test or per-fold scatter is provided. Thus, even setting aside the dataset-provenance issue, the headline improvement is not statistically established. The efficiency contribution (18.6M parameters, 50 epochs) remains plausible, but the specific claim of outperforming an ensemble is not supported by the evidence as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes integrating Axial-Coronal-Sagittal (ACS) convolutions into the nnU-Net framework for 3D brain tumor segmentation, with two strategies for transferring 2D ImageNet pre-trained weights (ACS-ResNet18 and ACS-All), and a joint classification-segmentation (JCS) model using a frozen HGG/LGG classification encoder. The methods are trained for 50 epochs on a single fold of BraTS2018 and BraTS2020 validation data, and compared against published 5-fold cross-validation ensembles from nnU-Net baselines. The authors report reduced parameters (18.6M vs 88.6M) and claim comparable or superior performance to the ensembles, with the best BraTS2018 result (ACS-ResNet18-All+BN+DA) reaching 0.85665 Dice.","tokens_in":10569,"tokens_out":5434,"duration_ms":43811,"significance":"If the claims were fully supported, the paper would be a useful contribution to efficient medical image segmentation: it demonstrates a large parameter reduction and shows that transfer learning from 2D ImageNet weights can be adapted to 3D convolutions. The experiments report mean and standard deviation over runs, which is good practice. However, the central comparison to published ensemble numbers is not statistically grounded, and the cited baseline for BraTS2018 appears to be from the BraTS 2017 challenge, so the 'outperform ensemble' claim is currently not established. The efficiency story (fewer parameters, fewer epochs) is plausible and interesting, but needs stronger baseline evidence.","major_comments":[{"comment":"The baseline labeled 'CV [12]' in Tables 1 and 3 is cited to reference [12], which is the 'Contribution to the BraTS 2017 Challenge' paper, not a BraTS 2018 validation result. The text in Section 5.1 refers to 'Isensee et al.'s ensemble attempt in the BraTS2018 challenge,' but the cited source does not support that description. If the quoted numbers (0.85600 Dice, 5.25667 HD95) were produced on BraTS2017 data or under a different protocol, the comparison is invalid. The authors must verify the provenance of these numbers and, if necessary, rerun a BraTS2018 baseline under the same protocol or replace the reference with a genuine BraTS2018 comparison.","section":"Section 5.1, Tables 1 and 3"},{"comment":"The difference between the proposed ACS-ResNet18-All+BN+DA (0.85665) and the quoted CV [12] (0.85600) is 0.00065 Dice, which is smaller than the reported standard deviation of the proposed method (0.00194). No statistical significance test, confidence interval, or per-case analysis is provided, so the claim that the method 'outperforms' the ensemble is not supported. Moreover, Table 6 shows that on BraTS2020 the method (0.85546) is slightly below the CV ensemble (0.85580), so the wording 'comparable or even outperform' overstates the evidence. The authors should report significance tests or per-fold results and temper the claim accordingly.","section":"Section 5.1, Table 3"},{"comment":"Experiments are performed only on the first fold of nnU-Net's 5-fold split, while the comparison baselines are 5-fold ensembles. A single fold may not be representative, and choosing the first fold without justification risks an optimistic comparison. The authors should report results across all folds (or at least the mean and range over folds) for their methods, or alternatively compare against a single-fold baseline trained under identical conditions. Without this, the robustness of the efficiency claim is unclear.","section":"Section 4.1 and all result tables"}],"minor_comments":[{"comment":"The column header 'SD Dice↓' should read 'SD HD95↓'.","section":"Table 2 header"},{"comment":"The sentence 'We include the result of Isensee et al. on the BraTS2018 validation set in row four of Table 2' is confusing because the CV [12] row appears in Table 1, not Table 2; please correct the cross-reference.","section":"Section 5.1"},{"comment":"Equation (2) defines the standard Hausdorff distance, not the 95th percentile version described in the text; please clarify how HD95 is computed.","section":"Section 4.2, Eq. (2)"},{"comment":"ACS-ResNet18-All+BN+DA+Post reports the same HD95 (4.87734, SD 0.24654) as ACS-ResNet18-All+BN+DA in Table 3, which suggests a possible copy-paste error; please verify.","section":"Table 5"},{"comment":"The slicing/reshaping of pre-trained weights (e.g., taking the first 320 channels from a 512-channel layer) is described without ablations comparing it to alternatives; an ablation of the channel-selection strategy would strengthen the paper.","section":"Section 3.1"},{"comment":"The phrase 'Four floating point precision' should read 'four decimal places'.","section":"Table 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant and timely problem, and the efficiency gains are potentially interesting. The main weakness is that the central claim of outperforming an ensemble rests on a questionable baseline citation and a negligible, statistically untested difference. I recommend major revision with a strong request to either rerun baselines under the same protocol or substantially soften the claims and provide per-fold or significance evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"If you care about cheap initialization for 3D segmentation, this paper is worth a look: they get an 18.6M-parameter nnU-Net variant to train in 50 epochs and land within a few thousandths of Dice of published 5-fold ensembles. That is a real practical result, and the two weight-transfer heuristics (matching-layer and sliced full transfer) are described concretely enough to reproduce. The JCS adapter idea is interesting, though it underperforms the ACS variants; the authors say so plainly, which I respect.\n\nThe soft spots are in the headline comparison. The row labeled 'CV [12]' in the BraTS2018 tables is from the BraTS 2017 challenge paper, not BraTS2018, and the text explicitly claims it is a BraTS2018 validation result. If those numbers came from a different split or protocol, the comparison is invalid. On BraTS2020, the same method is slightly below the quoted ensemble (0.85546 vs 0.85580), so the abstract's 'outperform ensemble' is simply not supported by the valid dataset. And on BraTS2018 the gap of 0.00065 Dice is smaller than the reported standard deviation of 0.00194; there is no significance test and no per-fold scatter. The single-fold setup only makes this worse.\n\nThe efficiency story also has a wrinkle: the ACS variant has 5x fewer parameters but is actually slower at inference (6.43s vs 4.78s), because ACS convolutions lack native implementation. That tempers the 'efficient' framing if runtime matters.\n\nThe slicing of ImageNet weights into 4-channel and 320-channel ACS convolutions is the load-bearing trick, but it is hand-specified and never ablated. I would like to see a random-init control and a study of the slicing choice. The lack of code also makes it hard to trust the numbers fully.\n\nWho is this for? Someone working on transfer learning for 3D medical segmentation, especially if they care about reducing training epochs. The paper deserves a serious referee, because the engineering is plausible and the experiments are carefully reported, but it needs major revision: rerun the baselines under the same protocol or drop the ensemble claim, add statistical tests and all-fold results, release code, and ablate the slicing decisions.\n\nMy recommendation: send it to peer review, but expect the authors to do significant work before it is publishable.","headline":"Plausible efficiency win, but the 'outperform ensemble' claim rests on a mislabeled baseline and a sub-SD gap.","tokens_in":11156,"tokens_out":2126,"would_cite":false,"duration_ms":20276,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A 50-epoch, single-fold nnU-Net with ImageNet-initialized ACS convolutions matches a 5-fold ensemble on brain tumor segmentation.","keywords":["brain tumor segmentation","nnU-Net","ACS convolutions","transfer learning","ImageNet pre-training","joint classification and segmentation","BraTS","Dice score"],"falsifier":"Compute the per-fold Dice of nnU-Net's 5-fold split on BraTS2018 and average over folds, or rerun the CV model on exactly the first fold for 50 epochs; if the first fold's mean Dice is notably above the 5-fold average, the 'outperform ensemble' claim collapses. Alternatively, train ACS-ResNet18-All+BN+DA with weights sliced from a different channel subset or with random initialization; if performance drops only slightly, the ImageNet transfer is not the operative ingredient.","tokens_in":10065,"feed_emoji":"🧠","tokens_out":4876,"duration_ms":40079,"temperature":0.7,"pith_summary":"The paper tries to establish that expensive 5-fold ensemble training is not necessary for state-of-the-art brain tumor segmentation: a single nnU-Net variant trained for only 50 epochs on one fold, with ACS convolutions initialized from ImageNet-pretrained ResNet18 weights, reaches or slightly exceeds the published ensemble Dice on BraTS2018 and comes within 0.0003 Dice of the BraTS2020 cross-validation ensemble. The practical stake is a roughly five-fold reduction in trainable parameters (18.6M vs 88.6M) and a twenty-fold reduction in epochs, which would make high-accuracy segmentation feasible with a single GPU. The paper also introduces a joint classification-and-segmentation model as a second route to improving challenging tumor labels.","feed_headline":"Single fast model matches 5-model brain tumor ensemble","feed_subtitle":"ImageNet-initialized ACS convolutions cut parameters 5x and epochs 20x while matching segmentation accuracy.","key_machinery":"Axial-Coronal-Sagittal (ACS) convolutions, which factor a 3D convolution into three 2D convolutions applied along the three anatomical planes, serve as the bridge for transferring 2D ImageNet weights into the 3D nnU-Net encoder. Two transfer strategies are used: ACS-ResNet18 initializes only layers whose input/output channels match ResNet18, while ACS-All slices and reshapes connected ResNet18 layers (e.g., the first three 64x64 layers) to cover nnU-Net's 4x32, 32x32, and 32x64 stages. A third variant, JCS, freezes a separately trained HGG/LGG classification encoder and fuses its features into the segmentation encoder through six JCS adapters, each a concatenation followed by a Squeeze-and-Excitation block and a 3x3 convolution.","core_discovery":"On BraTS2018 validation, ACS-ResNet18-All+BN+DA reports mean Dice 0.85665 and HD95 4.87734, outperforming the quoted 5-fold ensemble of the BraTS 2017 contribution (0.85600 Dice, 5.25667 HD95) despite training for 50 epochs on a single fold. With empirical post-processing, ACS-ResNet18+BN+DA+Post reaches 0.86231 Dice. On BraTS2020, ACS-ResNet18+BN+DA+Post attains 0.85546 Dice and 11.44526 HD95, close to the winning CV ensemble's 0.85580 Dice and better HD95. The authors interpret these results as evidence that 2D ImageNet-pretrained weights, transferred through ACS convolutions, can replace random initialization and lengthy multi-model ensembling.","pith_inferences":["The decisive comparison to a 5-fold ensemble rests on quoted published numbers; a fair test would rerun the CV model on the same first-fold split and post-processing, since the first fold could be easier than the average fold.","The paper does not ablate the slicing choices (e.g., first-320-channels versus other selections), so the degree to which transfer learning, rather than the architecture change, drives the gain is untested; a random-init ACS-ResNet18-All baseline would isolate this.","If the result transfers to later BraTS datasets, the practical protocol for surgical planning could become: take one fold, 50 epochs, one GPU, then fix small tumor components by a voxel threshold.","The JCS model's underperformance on BraTS2020 (0.84291 Dice) suggests the proxy-task benefit is dataset- and label-dependent, so generalization of the classification-to-segmentation transfer remains open."],"forward_implications":["A single model trained for 50 epochs on one fold can substitute for a five-model cross-validation ensemble, cutting training cost by roughly two orders of magnitude in epoch-fold product.","ImageNet's 2D representations survive a hand-designed slicing into 4-channel and 320-channel 3D convolutions, so pre-training on natural images is a usable initialization for 3D medical segmentation.","Batch normalization plus BraTS-specific augmentation and a voxel-count post-processing threshold can recover HD95 performance without extra epochs.","The 18.6M-parameter ACS encoder is about one-fifth the size of the 88.6M-parameter nnU-Net baseline, which helps deployment on memory-limited hardware.","The JCS adapter mechanism offers a template for injecting classification knowledge into segmentation, and the paper reports it lifts the challenging enhancing-tumor Dice to 0.8016."],"supporting_citations":[{"why":"Supplies the published BraTS2018 5-fold CV ensemble (0.85600 Dice, 5.25667 HD95) that the paper claims to outperform.","marker":"[12]"},{"why":"Supplies the BraTS2020 winning nnU-Net CV ensemble (0.85580 Dice) used as the comparison baseline on that dataset.","marker":"[11]"},{"why":"Provides the nnU-Net framework, planning files, and default training scheme that all variants build on.","marker":"[10]"},{"why":"ResNet18 is the source of the ImageNet-pretrained weights transferred into the ACS encoder.","marker":"[8]"},{"why":"ImageNet is the large-scale natural image dataset from which the pre-trained weights originate.","marker":"[5]"},{"why":"Introduces ACS convolutions, the mechanism that allows 2D pre-trained weights to be applied to 3D volumes.","marker":"[23]"},{"why":"Inspires the JCS adapter design that fuses frozen classification features into the segmentation encoder.","marker":"[22]"},{"why":"Provides the region-based training scheme used by default across all model variants.","marker":"[13]"}],"fun_headline_variants":["One model beats 5-model tumor segmentation ensemble","Fast single-model tumor segmentation rivals 5-fold ensemble","ACS nets cut training 20x, match tumor ensembles","ImageNet-pretrained ACS model matches tumor CV ensemble"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quoted 5-fold ensemble scores were obtained under a different training protocol, and the paper assumes those published numbers are directly comparable to its own single-fold, 50-epoch results without rerunning the baselines under the same fold, augmentation, and post-processing.","fun_headline_variants_meta":{"raw":{"variants":["One model beats 5-model tumor segmentation ensemble","Fast single-model tumor segmentation rivals 5-fold ensemble","ACS nets cut training 20x, match tumor ensembles","ImageNet-pretrained ACS model matches tumor CV ensemble"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1338,"prompt_tokens":908,"completion_tokens":430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":366}},"tokens_in":524,"tokens_out":430,"duration_ms":4525,"temperature":1.0,"reasoning_tokens":366,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:04:54.131244+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the per-fold Dice of nnU-Net's 5-fold split on BraTS2018 and average over folds, or rerun the CV model on exactly the first fold for 50 epochs; if the first fold's mean Dice is notably above the 5-fold average, the 'outperform ensemble' claim collapses. Alternatively, train ACS-ResNet18-All+BN+DA with weights sliced from a different channel subset or with random initialization; if performance drops only slightly, the ImageNet transfer is not the operative ingredient.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the published BraTS2018 5-fold CV ensemble (0.85600 Dice, 5.25667 HD95) that the paper claims to outperform."},{"cited_title":"In: Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 6th International Workshop, BrainLes 2020, Held in Conjunction with MICCAI 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the BraTS2020 winning nnU-Net CV ensemble (0.85580 Dice) used as the comparison baseline on that dataset."},{"cited_title":"In: 2009 IEEE CVPR","cited_arxiv_id":null,"evidence_quote":"ImageNet is the large-scale natural image dataset from which the pre-trained weights originate."},{"cited_title":"In: MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part I 24","cited_arxiv_id":null,"evidence_quote":"Introduces ACS convolutions, the mechanism that allows 2D pre-trained weights to be applied to 3D volumes."},{"cited_title":"IEEE TIP30, 3113–3126 (2021)","cited_arxiv_id":null,"evidence_quote":"Inspires the JCS adapter design that fuses frozen classification features into the segmentation encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the region-based training scheme used by default across all model variants."}],"review_version":1}