{"id":"b77cac28-abc6-4b01-87fd-dd8215b82f05","arxiv_id":"2603.00127","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An unsupervised self-annotation method using SLIC superpixels and a U-Net improves aggregate/mortar segmentation in low-contrast concrete XCT images compared to direct thresholding.","lead":"This paper applies a self-annotation-based unsupervised segmentation method to X-ray CT images of concrete, where aggregates and mortar look nearly identical. It shows that a U-Net model trained with this approach identifies aggregates better than direct greyscale thresholding on a held-out sample.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported advantage over direct thresholding may reflect per-image manual threshold tuning and unequal postprocessing rather than the self-annotation learning.","rationale":"The reader's weakest_assumption concerns superpixel misalignment in the self-annotation loop, which is a legitimate mechanistic risk. However, the single most load-bearing condition for the central claim—that the method outperforms direct thresholding—is that the evaluation comparison is fair and reproducible. The paper itself reveals that thresholds are per-image and manually chosen, and that the two compared pipelines use different postprocessing. This directly threatens the quantitative evidence for the abstract's claim, and it can be settled by a re-analysis with fixed thresholds and matched morphology. I therefore partially agree with the reader: the superpixel assumption is important, but the thresholding/evaluation protocol is the more immediate concern. The recommended verdict remains conditional acceptance: the method is promising, with code and data released, but the headline comparison must be re-validated under a pre-registered threshold rule before the claim can be taken at face value. No ad hominem is intended; the issue is with the evidence, not the authors.","tokens_in":24267,"tokens_out":8158,"duration_ms":84739,"concrete_test":"Using the released Zenodo code and data, recompute the aggregate metrics on the 16 manually labeled slices with thresholds chosen by a fixed automatic rule (e.g., Otsu) applied separately to the SS3 aggregate logit channel and to the raw TPR intensity, and with identical morphological postprocessing for both pipelines (same hole-filling, erosion, object-size cutoffs). Report per-slice IoU, F1, precision, and sensitivity, and repeat for all 10 training runs to bound run-to-run variability. If the logit-thresholding advantage shrinks or disappears under this fair comparison, the central claim is an artifact of threshold tuning and postprocessing differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim—that the self-annotation method 'consistently outperformed direct greyscale thresholding' for aggregate identification—rests on a comparison that is not yet controlled. In Section 6, the threshold applied to the aggregate channel of the normalized SS3 output is described as 'unique' for each image, but no selection rule is given. Section 5.3 explicitly omits thresholding details ('specific details pertaining to thresholding are not presented here'). Meanwhile, Section 7 shows that the logit-thresholding pipeline uses a more aggressive postprocessing chain (hole filling to 512 px, binary erosion, removal of objects <128 px, final dilation) than the direct-thresholding baseline (only removal of objects <128 px and holes <256 px). These differences confound the comparison: the reported gains in IoU, precision, sensitivity, and F1 could arise from manual per-image threshold tuning of the logit channel and from the extra morphological operations, rather than from the model having learned a genuinely better aggregate/mortar separation. Since the paper's practical contribution is precisely that the trained logit channel discerns aggregates better than raw intensity, this is the most load-bearing weak point: until the thresholds are fixed by a reproducible rule and the postprocessing is matched, the headline empirical claim is not substantiated. The self-annotation assumption (Section 3.4) is also a real risk, but the evaluation protocol is a more direct threat to the stated comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a self-annotation-based segmentation pipeline for low-contrast X-ray CT images of concrete, targeting three phases: aggregate, mortar, and porosity. A U-Net is trained without manually labeled aggregate/mortar data; per-pixel predictions are converted into 'dynamic image labels' by per-channel normalization, argmax classification, and SLIC superpixel refinement (Eq. 12, Section 3.4). Three training configurations are tested: fully unsupervised with 3 channels (US3), fully unsupervised with 4 channels (US4), and semi-supervised with 3 channels where porosity labels are provided by thresholding (SS3). On a held-out XCT sample with 16 manually annotated slices, the authors compare aggregate segmentation obtained by thresholding the aggregate-specific logit channel of SS3 against direct grayscale thresholding of the original XCT images, reporting that the proposed method consistently outperforms direct thresholding across IoU, precision, sensitivity, and F1. The authors also document limitations, including failure of the fully unsupervised variants to resolve porosity and systematic errors near sample peripheries and between closely spaced aggregates.","tokens_in":24569,"tokens_out":3032,"duration_ms":30448,"significance":"If the empirical claim is substantiated, the work has practical value: it offers a way to train a segmentation model for concrete XCT without manually annotated aggregate/mortar labels, which is a real bottleneck for morphology-based analysis in experimental mechanics. The paper also ships processed data, code, and model weights via Zenodo, which supports reproducibility and follow-up work. The held-out manual ground truth provides a meaningful independent check on the self-annotation loop. However, the headline claim is currently under-supported because the comparison against direct thresholding is not controlled, and the only fully successful configuration (SS3) is semi-supervised rather than unsupervised. The contribution of the self-annotation mechanism itself, as distinct from per-image threshold tuning and heavier morphological postprocessing, is not yet isolated.","major_comments":[{"comment":"The central claim that logit thresholding of the SS3 output 'consistently outperformed' direct grayscale thresholding is confounded. Section 6 states that the threshold applied to the aggregate channel is 'unique' for each image but gives no selection rule, and Section 5.3 explicitly omits thresholding details. Direct thresholding also appears to be manually tuned. Furthermore, the postprocessing chains differ: the logit pipeline applies hole filling to 512 px, binary erosion, removal of objects <128 px, and final dilation, while the direct-threshold baseline only removes objects <128 px and holes <256 px. Reported gains could therefore arise from per-image threshold selection and extra morphological operations rather than from the learned aggregate/mortar separation. Please fix both operations to a reproducible rule (e.g., Otsu or a fixed percentile on each channel) and use identical po","section":"§7.2, Fig. 31; §6; §5.3"},{"comment":"The paper is framed as an 'unsupervised approach,' but the fully unsupervised variants US3 and US4 both fail to unambiguously segment the porous phase (Figs. 21–24). The successful SS3 configuration uses thresholded porosity labels during training, making it semi-supervised. The abstract's claim that the methodology consistently outperforms direct thresholding refers to SS3, not to the unsupervised variants. This is a load-bearing mismatch between the title/abstract and the experimental results. Either temper the 'unsupervised' claims to 'semi-supervised for porosity, unsupervised for aggregate/mortar,' or demonstrate a fully unsupervised configuration that resolves all three phases.","section":"§5.1–§5.3; Abstract"},{"comment":"The self-annotation mechanism rests on the assumption that the most frequent predicted class within a SLIC superpixel is the correct semantic label for all pixels in that superpixel. If superpixels cross aggregate/mortar boundaries—which is likely in low-contrast regions—incorrect labels are fed back as training targets. The assumption is stated but never validated, e.g., by comparing superpixel boundaries against the manually annotated test slices or by measuring boundary recall/precision. A quantitative sanity check on the 16 annotated slices, or a sensitivity analysis over SLIC compactness and superpixel count, would materially strengthen the claim that the self-annotation loop is learning meaningful phase boundaries rather than propagating its own errors.","section":"§3.4, Eq. (12)"}],"minor_comments":[{"comment":"Typo: 'numcerical' should be 'numerical.' Also, the description of beam-hardening correction would benefit from a reference or more detail on how the manual parameter choice affects downstream slices.","section":"§2.1"},{"comment":"The loss curves are averaged over 10 training runs, but the qualitative results are shown for a single unspecified run. Please state clearly whether the presented segmentation examples are representative of all runs and whether the quantitative evaluation in Fig. 31 uses one model or an ensemble/average.","section":"§5, Fig. 20"},{"comment":"Only aggregate-phase metrics are reported. While mortar is complementary to aggregate, porosity is not evaluated quantitatively. Since porosity is a key phase in concrete morphology and SS3 is the only configuration that resolves it, reporting porosity IoU on the same ground truth would make the evaluation more complete.","section":"§7.2"},{"comment":"The normalization in Eq. (10) uses the per-image mean and variance of each channel, but the notation does not make clear whether these statistics are computed per tile, per image, or per batch. Please clarify, as this affects the interpretation of the 'dynamic labels' in minibatch training.","section":"Eq. (10)"},{"comment":"The suggestion to interpolate thresholds across slices using PCHIP is useful, but the manuscript does not state how many slices are needed for reliable interpolation or how sensitive the final metrics are to the choice of interpolation points. A short experiment or guidance would improve practical reproducibility.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a real problem and the open data/code is a plus. My recommendation is driven by the uncontrolled comparison in the central evaluation and the gap between the 'unsupervised' framing and the semi-supervised configuration that actually works. Both issues are fixable within the manuscript's scope: the thresholding/postprocessing comparison can be made fair, and the claims can be re-scoped. I do not see a fundamental flaw that would require rejection, but the current evidence does not yet support the headline claim as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on mesoscale concrete imaging. The real contribution is adapting Kanezaki's 2018 unsupervised segmentation to a U-Net on low-contrast concrete XCT, with SLIC superpixels generating dynamic labels. They ship code, data, and weights on Zenodo, which is a real plus. The paper is also honest: they clearly report that the fully unsupervised variants (US3, US4) fail to resolve the porosity phase, and only the semi-supervised SS3—using thresholded porosity labels—produces a clean three-phase output. Evaluation against manually annotated held-out slices is another genuine strength, and the ten-run loss curves suggest the training behavior is stable.\n\nThe soft spot is the central empirical claim. The paper says the method 'consistently outperformed direct greyscale thresholding' for aggregate identification, but the comparison is not controlled. Section 5.3 explicitly omits thresholding details, Section 6 says the threshold for each image is unique, and Section 7 applies a heavier postprocessing chain to the logit channel (hole filling to 512 px, erosion, removal of objects <128 px, dilation) than to the direct-threshold baseline (removal of objects <128 px and holes <256 px). With per-image manual thresholds, part of the reported gain could easily come from tuning the logit channel to each test slice rather than from the learned representation. The stress-test note lands. Until the threshold is fixed by a reproducible rule and the postprocessing is matched, the superiority claim is not substantiated. There is also only one held-out sample and 16 annotated slices, with no error bars on the quantitative metrics across the ten runs.\n\nNo mathematical flaws; the citation pattern is fair—Kanezaki is cited squarely and the contribution is framed as an adaptation. The main risks are overclaiming the word 'unsupervised' and overstating the comparison. A serious referee could get this into good shape. I would want the threshold rule specified, the postprocessing matched, and ideally the original Kanezaki method included as a baseline.\n\nSend it to peer review, but expect meaningful revision before publication.","headline":"Useful, honest adaptation of Kanezaki-style self-annotation to low-contrast concrete XCT, but the headline claim against direct thresholding is not yet controlled because of manual per-image thresholds and asymmetric postprocessing.","tokens_in":25064,"tokens_out":2587,"would_cite":false,"duration_ms":26157,"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":"This paper shows that a U-Net trained with self-annotated superpixel labels can separate aggregate from mortar in low-contrast concrete XCT images, outperforming direct greyscale thresholding on out-of-distribution data.","keywords":["X-ray computed tomography","concrete","unsupervised semantic segmentation","self-annotation","superpixels","U-Net","aggregate segmentation","low-contrast XCT"],"falsifier":"Compare the superpixel-refined training labels against the 16 manually annotated slices: if superpixels frequently straddle aggregate-mortar boundaries in low-contrast regions and the model's errors correlate with those straddles, the refinement assumption fails. More directly, train the same pipeline with deliberately misaligned superpixels (e.g., extreme compactness or oversized superpixels); if aggregate identification accuracy does not degrade, the superpixel-alignment assumption is not load-bearing.","tokens_in":24113,"feed_emoji":"🧱","tokens_out":4849,"duration_ms":42350,"temperature":0.7,"pith_summary":"The paper tries to establish that low-contrast X-ray CT images of concrete—where aggregate and mortar look almost identical in intensity—can be segmented without any labelled training data, by having the model teach itself. The proposed loop uses a convolutional encoder-decoder network to produce rough per-pixel phase predictions, then SLIC superpixels to smooth those predictions into region labels by majority vote, and then trains the network on those self-generated labels as if they were ground truth. On a held-out concrete sample, the resulting aggregate maps beat direct greyscale thresholding on the metrics reported. This matters because concrete XCT segmentation normally requires costly manual annotation or contrast-enhancing additives, both of which limit practical use.","feed_headline":"Self-annotated U-Net beats direct thresholding on concrete XCT","feed_subtitle":"Superpixel labels let a CNN separate aggregates from mortar in low-contrast scans without manual annotation.","key_machinery":"The engine is the superpixel-refinement self-annotation loop. SLIC partitions each image into small contiguous regions of perceptual similarity; the most frequent predicted class in each region becomes the label for all its pixels, producing dynamic image labels that are fed back as training targets. This converts noisy, pixel-level CNN outputs into spatially coherent pseudo-labels and ties local superpixel structure to the global context captured by the network's receptive field. Per-channel normalization of the output prevents the model from collapsing to a single phase, although the paper notes it remains fragile under class imbalance. The architecture is a U-Net-style encoder-decoder wit","core_discovery":"The paper's central claim is that a convolutional encoder-decoder network can learn to separate aggregate from mortar in low-contrast concrete XCT images using no labelled data for those phases, and that the resulting aggregate segmentations are more accurate than direct greyscale thresholding. The self-annotation loop works as follows: after each training pass, per-pixel predictions are normalized per output channel, converted to hard labels by argmax, and then refined by replacing every pixel's label with the majority label of its SLIC superpixel. These refined labels—called dynamic image labels—are used as the cross-entropy training targets for the next iteration, evolving as the model im","pith_inferences":["A natural testable extension is to apply the same self-annotation loop to other low-contrast multiphase materials with a known phase count and check whether the superpixel-mode assumption degrades as phase-boundary contrast drops.","The reported periphery failure may be a receptive-field or padding artifact rather than a fundamental limit; training on full slices or using border-aware crops could recover those aggregates.","Because the paper flags per-channel normalization as fragile under class imbalance, a class-balanced or entropy-regularized alternative could make the fully unsupervised three-phase case succeed and remove the need to threshold porosity.","The dynamic-label loop is a form of self-training, so its convergence likely depends on initial random predictions being slightly better than chance within each superpixel; this could be tested by seeding training with a simple intensity prior."],"forward_implications":["Concrete XCT scans can be segmented into aggregate and mortar phases without manual annotation, removing the main bottleneck to using XCT in concrete mechanics.","Thresholding the aggregate-specific channel of the normalized network output identifies aggregates better than thresholding raw XCT slices, so the network acts as an intensity transform that sharpens phase separability.","A purely unsupervised three-phase setup fails to isolate porosity; the semi-supervised variant, which thresholds only the easily separable pore phase, gives unambiguous three-phase labels.","The learned segmentation can bootstrap supervised training: the paper explicitly proposes using its outputs as initial labels for fine-tuning on new concrete datasets.","The method inherits concrete failure modes: aggregates near the sample periphery are missed and closely spaced small aggregates are merged, so downstream users should expect those behaviors."],"fun_headline_variants":["Superpixel labels teach CNN to beat thresholding on concrete scans","Unsupervised CNN beats thresholding on low-contrast concrete CT","Self-annotation with superpixels improves concrete XCT segmentation","No labels needed: CNN wins over thresholding on concrete XCT","Superpixel-guided CNN segments concrete without labeled data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that SLIC superpixels stay inside single phases and that the majority predicted class within a superpixel is the true phase; in low-contrast regions where superpixels straddle aggregate-mortar boundaries, wrong labels are fed back into training and the model can learn to reproduce those errors.","fun_headline_variants_meta":{"raw":{"variants":["Superpixel labels teach CNN to beat thresholding on concrete scans","Unsupervised CNN beats thresholding on low-contrast concrete CT","Self-annotation with superpixels improves concrete XCT segmentation","No labels needed: CNN wins over thresholding on concrete XCT","Superpixel-guided CNN segments concrete without labeled data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3570,"prompt_tokens":794,"completion_tokens":2776,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":2700}},"tokens_in":538,"tokens_out":2776,"duration_ms":16151,"temperature":1.0,"reasoning_tokens":2700,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:31:29.597721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the superpixel-refined training labels against the 16 manually annotated slices: if superpixels frequently straddle aggregate-mortar boundaries in low-contrast regions and the model's errors correlate with those straddles, the refinement assumption fails. More directly, train the same pipeline with deliberately misaligned superpixels (e.g., extreme compactness or oversized superpixels); if aggregate identification accuracy does not degrade, the superpixel-alignment assumption is not load-bearing.","supporting_citations":[],"review_version":1}