{"id":"60918667-31c9-410b-a488-420d3b2c004e","arxiv_id":"2607.14287","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Sequential two-stage adaptation of SAM through an alloy-microstructure intermediate domain improves defect segmentation on synthetic and real X-ray CT benchmarks while training only 0.647% of parameters.","lead":"This paper tests a two-stage recipe for adapting the Segment Anything Model to find tiny defects in X-ray CT scans of 3D-printed metal parts. It first adapts SAM on alloy-microstructure photos, then on synthetic X-ray defect data, and reports the best pixel-level accuracy among the compared methods while training only 0.647% of the model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The two-stage benefit is confounded with total training compute: no direct Conv-LoRA-SAM baseline is trained for the same number of epochs/schedule as XCT-SAM, so the alloy stage may not cause the reported gains.","rationale":"The reader conditionally accepts the paper, and I agree with that verdict. The strongest claim includes attribution to sequential domain bridging, and that attribution is the novel component of the method. The paper's evidence is a comparison with a directly adapted Conv-LoRA-SAM baseline, but the compute budget and optimization schedule are not controlled. The implementation details specify only XCT-SAM's two-stage schedule, not the baseline's epoch count or LR schedule. Since LoRA/Conv-LoRA adapters are small, longer training often improves convergence, and warm-starting from alloy weights may help even if the alloy images are not semantically close to XCT. The paper's t-SNE visualization (Fig. 4) does not include the alloy dataset, so no feature-distance evidence demonstrates that alloy is an intermediate domain between natural images and XCT. A same-compute control is necessary and feasible. Secondary concerns, such as threshold-derived NIST pseudo-labels and small margins on real data, are present but are secondary to the causal attribution issue. No change to the reader's verdict is needed; it should remain CONDITIONAL until the same-compute control is added.","tokens_in":11259,"tokens_out":4539,"duration_ms":49158,"concrete_test":"Train three configurations with identical total steps, batch size, optimizer, and seeds on the GAN-XCT training data: (1) XCT-SAM as reported (15 alloy epochs + 20 XCT); (2) direct Conv-LoRA-SAM for 35 XCT epochs with the same LR schedule as XCT-SAM (2e-4 for first 15 epochs, then 8e-5 for 20); (3) direct Conv-LoRA-SAM for 35 XCT epochs at constant 8e-5. Evaluate on the same GAN and NIST sets. If condition (2) or (3) matches or beats XCT-SAM's Table I IoU/Dice/F1 margins, the alloy stage is not necessary; report mean and std over at least 3 seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adapting on alloy-microstructure data before XCT causes the gains. The only evidence for this is the comparison to Conv-LoRA-SAM in Table I, which is described as fine-tuned on GAN XCT without the intermediate alloy stage (Sec. IV-B). But Sec. IV-D gives XCT-SAM 15 epochs on alloy plus 20 epochs on XCT, with learning rates 2e-4 then 8e-5. The paper never states the Conv-LoRA-SAM baseline's epoch count, total steps, or LR schedule. If the baseline was trained only for the Stage-2 budget (20 epochs at 8e-5), then XCT-SAM's improvement could come from 35 total epochs, the higher initial LR, or the warm-start initialization, rather than from the alloy domain specifically. Table I's margins on pores are +0.0472 IoU, and on inclusions +0.1232 IoU; these are large enough to be explained by extra training. This is not an internal inconsistency, but it leaves the causal attribution unsupported. The ablation study (Tables II-III) varies rank and loss but never removes Stage 1 while holding compute fixed, so the sequential-bridging claim lacks its required control.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes XCT-SAM, a two-stage parameter-efficient domain adaptation method for SAM-based defect segmentation in additive-manufacturing XCT images. In Stage 1, Conv-LoRA adapters (rank r=2, eight convolutional experts) are fine-tuned on an alloy-microstructure dataset; in Stage 2, the adapters are further fine-tuned on CycleGAN-generated synthetic XCT data. The authors report evaluations on synthetic GAN-XCT benchmarks and real NIST XCT scans, comparing against zero-shot SAM, UNet++, MedSAM, SAM-Med2D, and direct Conv-LoRA-SAM. They report the best overall IoU and Dice across all settings with 4.15M trainable parameters (0.647% of the model), and release the code.","tokens_in":11628,"tokens_out":4187,"duration_ms":47203,"significance":"If the two-stage bridging claim is supported, the contribution is practically useful: it combines an intermediate material-domain adaptation stage with parameter-efficient Conv-LoRA to improve SAM on sparse, non-semantic industrial defects, and includes evaluations on an external real-world NIST dataset. Strengths include the public code release, the consideration of real XCT data, and ablations over rank and loss objectives. The main reservation is that the central attribution — that the intermediate alloy stage, not extra training compute or a different schedule, drives the gains — is not directly tested because no matched-compute direct baseline is included.","major_comments":[{"comment":"The central claim that the alloy intermediate stage causes the reported gains is not isolated from training compute and schedule. XCT-SAM is trained for 15 alloy epochs at LR 2e-4 plus 20 XCT epochs at LR 8e-5 (Sec. IV-D), while the direct Conv-LoRA-SAM baseline is described only as fine-tuned on GAN XCT without the alloy stage (Sec. IV-B); its epoch count, total updates, and LR schedule are not reported. If the baseline is trained only for the Stage-2 budget, the +0.0472 IoU (pores) and +0.1232 IoU (inclusions) margins in Table I could come from longer training, the higher initial LR, or warm-start initialization rather than from the alloy domain specifically. The ablation study in Tables II–III also never removes Stage 1 while holding compute fixed. I request a matched-compute control: train direct Conv-LoRA-SAM on XCT for the same total number of updates and with the same LR schedule","section":"Sec. IV-B and IV-D, Table I"},{"comment":"The evaluation protocol appears to include the validation set in the reported averages. Section IV-A states that Test-3 is used as the validation set because it clusters closely with Train-1 and Train-2, but Table I reports means computed 'across all test sets,' and the quantitative text in Sec. IV-E refers to 'all test sets' without excluding Test-3. Including the validation set in final averaged numbers can inflate reported performance and makes the benchmark less clean. Please report per-test-set results and recalculate the averages either excluding Test-3 or explicitly treating it as a held-out validation set with no hyperparameter selection.","section":"Sec. IV-A and Table I"},{"comment":"The NIST ground-truth masks are not publicly available, so the authors generate threshold-derived reference masks following the pipeline of [33]. This means the NIST results evaluate agreement with a threshold-based proxy rather than with expert annotations. The threshold value and preprocessing steps are not reported, and no sensitivity analysis is provided. Since the NIST benchmark is a major part of the claim of real-world generalization, the threshold-dependence of the reference masks should be quantified (e.g., report performance over a range of thresholds or compare against any available published segmentations). Without this, the NIST IoU/Dice numbers are difficult to interpret and reproduce.","section":"Sec. IV-A, NIST evaluation"}],"minor_comments":[{"comment":"The t-SNE analysis uses ViT-B/16 features, but the adapted model uses a ViT-H encoder. Please clarify whether the ViT-B/16 features are from the original SAM, from a Conv-LoRA-adapted model, and why this feature space is representative of the ViT-H backbone used in all experiments.","section":"Fig. 4"},{"comment":"The gating function and convolutional expert details are not fully specified. Please define the gating network, the kernel sizes/strides of the convolutional experts, and how the number of experts (eight) is chosen; this would also make the rank/parameter-count ablation in Table II more reproducible.","section":"Eq. (1)"},{"comment":"Table I reports only mean values without variance or statistical significance. Fig. 5 shows error bars, but the number of runs/seeds and the source of the variance are not stated in the table or text. Reporting per-test-set values or standard deviations in the table would make the margins between XCT-SAM and Conv-LoRA-SAM more interpretable.","section":"Table I and Fig. 5"},{"comment":"The F1 column appears to be the tolerance-based F1 defined in Sec. IV-C, but the table header uses 'F1' without the qualifier. Please rename it to 'Tolerance-F1' for consistency and to avoid confusion with the Dice coefficient.","section":"Table III"},{"comment":"The ablation studies vary rank and loss but do not vary the number of convolutional experts, although the abstract and conclusion highlight 'eight convolutional experts' as part of the best configuration. A brief sentence justifying this fixed value or an ablation over the number of experts would strengthen the parameter-efficiency analysis.","section":"Sec. IV-F"},{"comment":"The limitations paragraph mentions the additional cost of two-stage training but does not report the actual compute overhead (epochs, GPU-hours, or wall-clock time) relative to direct Conv-LoRA-SAM. Please include a quantitative comparison of training cost.","section":"Sec. V"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the empirical results are interesting, but the central causal claim — that sequential alloy-to-XCT adaptation is responsible for the gains — is not yet supported because the direct baseline is not matched in compute or schedule. The validation-set inclusion in Table I should also be addressed, as it affects the interpretation of the headline numbers. These issues are fixable with additional experiments and revised reporting, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-written engineering paper with a genuinely new recipe—adapting SAM to XCT through an intermediate alloy-microstructure stage—and the results on both synthetic and real NIST data are consistent. The implementation details are transparent, and the parameter efficiency is real (4.15M trainable params, 0.647% of the model). The loss and rank ablations are useful. But the paper's central claim—that the alloy stage, rather than extra compute or schedule, drives the gains—is untested. The direct Conv-LoRA-SAM baseline is trained on XCT without the alloy stage, but we never get its epoch count, step count, or LR schedule. If it is trained for only the Stage-2 budget (20 epochs at 8e-5), then XCT-SAM's advantage could come from 35 total epochs, the higher initial LR, or the warm start itself. The +0.0472 IoU on pores and +0.1232 IoU on inclusions are large enough to be explained by additional training. Tables II and III vary rank and loss, but never remove Stage-1 while holding compute fixed. So the 'sequential domain bridging' hypothesis remains plausible and motivated, but unsupported by the current experiments. A matched-compute baseline (direct adaptation for 35 epochs with the same schedule, or a warm-start-from-SAM control) would settle it. Minor issues: Table I reports averages over all test sets while Test-3 is used for validation; if hyperparameters were selected on Test-3, that score should be excluded or reported separately. The NIST 'ground truth' is threshold-derived, so the real-data evaluation is softer than the wording suggests. Also, the inclusion margins on NIST are omitted because that dataset has only pores—fine, but the abstract's 'consistent' outperformance should be read with that caveat. None of this is fatal. The paper is honest about its limitations (two-stage cost, 2D slice-wise binary models, noise sensitivity). For a workshop paper at ICPR/MVI2, this is a reasonable contribution. Deserves a serious referee: the core idea is interesting and the experiments are mostly careful; the missing compute-matched control is a fixable gap, not a methodological collapse.","headline":"Useful empirical recipe, but the central causal claim about sequential bridging is not yet supported—the two-stage model gets extra training time and a different schedule than the direct baseline.","tokens_in":12071,"tokens_out":2417,"would_cite":false,"duration_ms":24845,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sequentially adapting SAM to alloy microstructures before XCT defect data improves pore and inclusion segmentation with only 4.15M trainable parameters.","keywords":["XCT defect segmentation","Segment Anything Model","parameter-efficient fine-tuning","Conv-LoRA","domain adaptation","class imbalance","additive manufacturing"],"falsifier":"Train the direct Conv-LoRA-SAM baseline on the same XCT data for exactly the combined number of epochs (15 alloy + 20 XCT) with the same two-stage learning-rate schedule, except skip the alloy stage and either keep the same total steps or use the first 15 epochs on XCT data. If this matched-cost baseline matches or exceeds XCT-SAM's IoU/Dice on the synthetic and real test sets, the sequential-bridging claim fails.","tokens_in":11205,"feed_emoji":"🩻","tokens_out":4780,"duration_ms":49916,"temperature":0.7,"pith_summary":"This paper proposes XCT-SAM, a two-stage parameter-efficient strategy for adapting the Segment Anything Model (SAM) to defect segmentation in X-ray computed tomography (XCT) images of additively manufactured parts. The key claim is that first fine-tuning lightweight Conv-LoRA adapters on alloy-microstructure images—which resemble XCT texture—and then transferring the adapted model to XCT data improves pore and inclusion segmentation compared with zero-shot SAM, direct fine-tuning, and other SAM-adaptation baselines. The authors report consistent gains on both synthetic GAN-generated XCT benchmarks and real-world XCT scan data, while training only about 4.15 million parameters (0.647% of the frozen ViT-H backbone). A sympathetic reader would care because AM defect segmentation is data-scarce, severely class-imbalanced, and suffers from large distribution shifts; the paper argues that a cheap intermediate domain can make foundation-model adaptation feasible.","feed_headline":"Two-stage domain adaptation lifts SAM on industrial XCT defects","feed_subtitle":"Pre-training adapters on alloy microstructures before XCT fine-tuning beats zero-shot and direct fine-tuning with only 0.65% of parameters t","key_machinery":"Conv-LoRA, a parameter-efficient adapter that injects a low-rank update plus a set of convolutional expert gates into each transformer block of SAM's frozen ViT-H encoder. The adapter computes an embedding update of the form W0x + WD(Σ gi(x) Ei(WEx)), where WE and WD are low-rank projections, Ei are convolutional experts, and gi are gating weights. XCT-SAM's central use of this mechanism is sequential: the adapter weights are first trained for up to 15 epochs on alloy-microstructure images, then re-trained with a lower learning rate for 20 more epochs on GAN-generated XCT images. The paper argues that this warm start decomposes the large natural-image-to-XCT shift into two smaller shifts.","core_discovery":"On its own terms, the paper establishes that a curriculum-like sequential adaptation of SAM's frozen image encoder—first on real alloy-microstructure images, then on synthetic XCT defect images—produces better out-of-distribution defect segmentation than adapting on XCT alone. The gains are attributed to the intermediate stage warming up the low-rank adapter weights toward metallic, low-contrast spatial features, so the limited XCT labels drive a smaller domain shift. Measured on IoU, Dice, and F1 across six synthetic test sets and multiple real-world scan sets, the best configuration (rank 2, eight convolutional experts, Dice-Focal loss) outperforms every baseline on both pore and inclusion","pith_inferences":["The paper does not rule out that the alloy stage simply provides extra training time or a favorable learning-rate schedule; a matched step-count control without the alloy stage would be needed to isolate the bridging effect.","The approach may extend to 3D volumetric segmentation by slotting the same two-stage adapter schedule into a volumetric SAM variant, though the paper tests only 2D slices.","Because the real-world reference masks are threshold-derived, part of the reported real-world gain could reflect alignment with the thresholding pipeline rather than with true defect geometry; ground-truth labels from human annotation or higher-resolution scans would be a stronger test.","The two independent binary models create redundant computation; a unified multiclass model with a class-balanced loss might reach similar accuracy at lower inference cost."],"forward_implications":["If the sequential adaptation claim holds, SAM-based industrial XCT defect segmentation becomes practical on small labeled datasets, with most of the model frozen.","The intermediate-domain warm-start recipe could be reused for other non-destructive-testing modalities where labeled defect data are scarce but related material-imaging data exist.","The rank-2 result suggests that very low-rank adapters are sufficient for sparse, low-contrast defect domains, and that higher ranks can hurt rare-class generalization.","The loss-function study indicates that Dice-Focal is a robust default for extreme class imbalance, outperforming plain Dice, BCE, Lovász-Softmax, and Focal Tversky on the tested benchmarks."],"fun_headline_variants":["Sequential SAM adaptation beats direct fine-tuning on XCT defects","Warm-start adapters on alloys to boost SAM on XCT defects","Intermediate alloy step sharpens SAM for industrial XCT","Tiny adapters, big gains: sequential SAM adaptation for XCT"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that the intermediate alloy training stage—rather than the extra training time or different optimization schedule it introduces—is what drives the performance improvement is not directly tested, because the direct baseline receives fewer total training updates.","fun_headline_variants_meta":{"raw":{"variants":["Sequential SAM adaptation beats direct fine-tuning on XCT defects","Warm-start adapters on alloys to boost SAM on XCT defects","Intermediate alloy step sharpens SAM for industrial XCT","Tiny adapters, big gains: sequential SAM adaptation for XCT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1078,"prompt_tokens":815,"completion_tokens":263,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":194}},"tokens_in":559,"tokens_out":263,"duration_ms":3810,"temperature":1.0,"reasoning_tokens":194,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T02:31:51.362972+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the direct Conv-LoRA-SAM baseline on the same XCT data for exactly the combined number of epochs (15 alloy + 20 XCT) with the same two-stage learning-rate schedule, except skip the alloy stage and either keep the same total steps or use the first 15 epochs on XCT data. If this matched-cost baseline matches or exceeds XCT-SAM's IoU/Dice on the synthetic and real test sets, the sequential-bridging claim fails.","supporting_citations":[],"review_version":1}