{"id":"dc9e1279-f91f-42d3-a7d9-93c34b64c0fb","arxiv_id":"2412.06499","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid CNN-Transformer network with bi-level routing attention and feature fusion reports state-of-the-art or near-state-of-the-art landmark detection accuracy across five medical X-ray datasets.","lead":"HYATT-Net is a deep learning model that finds anatomical landmarks, such as skull, hand, and pelvis reference points, in medical X-rays. It combines convolutional and transformer components with a sparse attention mechanism and reports small accuracy gains over prior state-of-the-art models on five datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on baseline comparisons whose protocol match is unverified; internal table errors (CeLDA SDR non-monotonic, ISBI2023 text citing ISBI2015 SDR) make the reported margins unreliable.","rationale":"I agree with the reader's weakest assumption. The concern is load-bearing because the paper's headline contributions are the SOTA margins; without protocol-matched baselines, the margins are not interpretable. The paper's internal inconsistencies (non-monotonic CeLDA SDR; wrong SDR text) make this more than a theoretical worry. A positive note: the paper states code and data are released, which makes the proposed reproduction test feasible; if the repository reproduces the table values and protocol, the accuracy claim would be substantially supported. However, because the central claim depends on unverified baseline comparability and the paper already contains data-integrity red flags, the verdict should remain CONDITIONAL, not downgraded. I do not see a reason to move to REJECT: the architecture is plausible, the method is clearly described, and the accuracy claim may survive a matched-protocol rerun.","tokens_in":17331,"tokens_out":6789,"duration_ms":69300,"concrete_test":"Using the released repository, reproduce the HTC and FARNet rows of Tables 1 and 2 under the exact HYATT-Net configs (same MMPose pipeline, resize, augmentation, split, and evaluation code) on ISBI2023 and Hand. If the reproduced baseline MREs differ from the published rows by more than the claimed HYATT-Net margins (0.03mm on ISBI2023, 0.03mm on Hand) or if any reproduced baseline falls below HYATT-Net's MRE, then the SOTA comparison is not protocol-matched and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is comparative: HYATT-Net beats HTC/FARNet/CeLDA by 0.03-0.07mm MRE (and 0.59px on pelvis). These margins are smaller than typical protocol differences, yet Tables 1-3 mix published numbers (starred) with unstarred in-house runs without stating which baselines were rerun, under which preprocessing, augmentation, split, and metric conventions. The concern is not merely hypothetical: Table 1's CELDA average row gives CeLDA SDR 99.67 at 3mm and 98.67 at 4mm, which violates monotonicity in the radius; and Section 4.4.1 attributes 84.78/89.92 SDR values (the ISBI2015 numbers) to ISBI2023, where the table shows 87.43/91.43. These inconsistencies indicate the table values cannot be treated as a reliable protocol-matched comparison. If the unstarred baselines were run by the authors using their own pipeline, the settings are not documented; if the starred numbers come from different papers, they may differ in image resizing (1024x1216 vs 1024x1024 vs 512x512), hand scaling convention (50mm), train/test splits, and landmark ordering. Because the claimed improvements are comparable to or smaller than the ablation effects (CBAM: 0.007mm; FFCM: 0.005mm) and inter-observer variability, protocol mismatch alone could erase the SOTA claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes HYATT-Net, a U-shaped hybrid CNN-Transformer architecture for anatomical landmark detection from X-ray images. The main components are a BiFormer module using bi-level routing attention for efficient global context, an Attention Residual Module with CBAM for local feature refinement, a Feature Fusion Correction Module that combines multi-scale features with global pooling, and deep supervision with MSE loss on multi-resolution heatmaps. The authors evaluate on five datasets (ISBI2015, ISBI2023, CephAdoAdu, Hand X-rays, Pelvic X-rays) and report MRE/SDR improvements over several baselines, e.g., MRE 1.13 mm on ISBI2015 versus 1.19 mm for HTC and 0.98 mm on CephAdoAdu versus 1.05 mm for CeLDA. Code and data are released on GitHub.","tokens_in":17642,"tokens_out":4991,"duration_ms":47209,"significance":"If the reported results are taken at face value and the baseline comparisons are protocol-matched, the work offers a modest but consistent improvement over prior state of the art across several anatomies, with released code that would benefit the community. The architecture itself is a combination of existing components (BiFormer, CBAM, deep supervision, feature fusion), so the novelty lies in the specific integration for ALD rather than in a new attention mechanism. The main strengths are the breadth of datasets, the public release of code/data, and the consistent direction of the reported improvements. However, the claimed margins over baselines (0.03–0.07 mm) are comparable to or smaller than the ablation differences within the paper and well below the reported standard deviations; the reliability of the quantitative claims is also undercut by text/table inconsistencies and missing protocol documentation. The paper is a plausible contribution if these issues are resolved.","major_comments":[{"comment":"The text states that on ISBI2023 'the SDR at 2.0mm and 2.5mm reaches 84.78% and 89.92%', but Table 1 reports 87.43% and 91.43% for ISBI2023; the 84.78/89.92 values are the ISBI2015 numbers. This mismatch makes it impossible for the reader to determine which dataset the quoted margins refer to, and it affects the claimed superiority over HTC and FARNet on ISBI2023. Please correct the text or the table and re-verify all quoted differences after the correction.","section":"Section 4.4.1, Table 1"},{"comment":"The reported CeLDA SDR values are 89.13% at 2 mm, 93.60% at 2.5 mm, 99.67% at 3 mm, and 98.67% at 4 mm. Since SDR is a cumulative success rate within a radius, it must be non-decreasing in the threshold; 99.67% at 3 mm and 98.67% at 4 mm is impossible and indicates an entry error or an inconsistent definition. Please correct this row and run the same monotonicity check on all SDR columns, as a non-monotonic SDR also appears in the CELDA (Teenager) row where 96.90% at 3 mm is followed by 98.33% at 4 mm (this is allowed) but the average row is not.","section":"Table 1, CELDA (Average) row"},{"comment":"The SOTA comparisons mix published numbers (marked with *) with apparently in-house numbers without stating which baselines were rerun and under what protocol. Differences in image resizing (1024×1216 vs 1024×1024 vs 512×512), augmentation, train/test splits, landmark sets, metric conventions, and the 50 mm hand-scaling assumption can plausibly shift MRE by more than the claimed 0.03–0.07 mm margins. Because the central claim is comparative, please document the exact protocol for every baseline or rerun all baselines under the same pipeline; otherwise the SOTA claim is not established.","section":"Section 4.2 and Tables 1–3"},{"comment":"The ablation attributes 0.007 mm improvement to CBAM and 0.005 mm to FFCM on ISBI2015. Given that the reported standard deviations are about 1.1–1.2 mm and no significance test or repeated-run statistics are provided, these differences are indistinguishable from noise. The claim that these modules 'further enhance SDR and reduce MRE' is therefore not supported by the evidence as presented. Please add variance across seeds, confidence intervals, or a paired significance test, or rephrase the attribution to avoid overclaiming.","section":"Section 4.5, Table 4"},{"comment":"The paper claims superiority in 'accuracy, robustness, and efficiency,' but Section 4 contains no efficiency measurements (parameters, FLOPs, inference time) and no explicit robustness experiments (e.g., noise, corruption, or domain shift beyond the five datasets). The multi-dataset results support generalizability, not the stated robustness and efficiency claims. Please either add the corresponding experiments and metrics or limit the claims to accuracy.","section":"Abstract and Section 4.4.3"}],"minor_comments":[{"comment":"The normalization factor in the 2D Gaussian heatmap is written as 1/(√(2π)σ), which is the 1D normalization; for a 2D isotropic Gaussian it should be 1/(2πσ²). As written the heatmap is not volume-normalized, although for training this may not matter in practice.","section":"Equation (1)"},{"comment":"The dimensions of Q, K, V and Q_p, K_p are inconsistent: after partitioning an H×W image into S×S patches, the number of patches is HW/S², so Q should be in R^{HW/S² × C} rather than R^{S² × HW/S² × C}. Please clarify the notation.","section":"Section 3.2"},{"comment":"Equation (4) defines the total loss with weights α₁ and α₂, and the text says α₁ = 3 and α₂ = 3; Section 4.3 instead reports w₁ = 1, w₂ = 3, w₃ = 3. Please reconcile these notations and values.","section":"Section 3.4 vs Section 4.3"},{"comment":"In the CephAdoAdu paragraph, the adolescent MRE is reported as 0.85 px, but the dataset and all other MRE values in that section are in mm; this is presumably a typo and should read 0.85 mm.","section":"Section 4.4.1"},{"comment":"The dataset is referred to inconsistently as 'CephAdoAdu' in the text and 'CELDA' in Table 1; please use a consistent naming convention and clarify whether 'CELDA (Average)' refers to the full CephAdoAdu test set.","section":"Table 1 and Section 4.1.1"},{"comment":"References [8] and [28] are the same paper (Lowe, 1999), and references [9] and [31] are the same paper (Liu et al., 2010); please merge the duplicates.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely acceptable after the protocol documentation and table corrections are made. Given that the claimed margins are small and the current tables contain impossible values, I suggest the editor ask the authors for corrected tables and a detailed baseline-protocol description before a second review round. The self-citation pattern is acceptable in this field, but the comparability of starred baseline numbers should be verified independently."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible but unproven SOTA claim wrapped around an incremental architecture. The paper does a lot right: five datasets, released code, consistent MRE improvements in their own tables, and a clear write-up of the hybrid design. The problem is the experimental evidence as presented doesn't yet support the 'surpassing existing methods' headline.\n\nWhat's new: the specific stack of BiFormer + CBAM-enhanced ARM + FFCM + deep supervision for ALD. Each piece is known, but the combination is new and a reasonable engineering choice. The FFCM is a small global-pooling-plus-FC module meant to recover resolution loss; deep supervision is standard.\n\nWhere it gets shaky: the SOTA comparison relies on mixing published numbers (starred) with the authors' own runs, and the paper never states which baselines were rerun under which preprocessing, split, augmentation, or metric convention. Given the claimed margins are 0.03–0.07mm MRE, protocol differences common in this literature could easily erase them. The stress-test note found concrete signs of trouble: Table 1 gives CeLDA an SDR of 99.67 at 3mm and 98.67 at 4mm, which is impossible for a cumulative success rate. The ISBI2023 paragraph cites SDR values (84.78/89.92) that are actually the ISBI2015 numbers. Those two errors mean the table values can't be taken at face value until corrected. The ablation is also weak: CBAM adds 0.007mm and FFCM 0.005mm, both far below the reported STD of ~1.1mm, so those components aren't shown to matter.\n\nThe central accuracy claim might still be true; the architecture is sensible and the published numbers for HTC and FARNet are roughly consistent with what others report. But the paper as written doesn't demonstrate it. A careful re-run with all baselines under the same protocol, plus corrected tables, would make this a solid contribution.\n\nBottom line: worth engaging with for a referee, but it needs major revision before I'd trust the SOTA numbers. For a reading group, it's a decent example of a common failure mode in applied medical vision comparisons.","headline":"A workmanlike hybrid architecture with a SOTA claim that isn't yet backed by the experiments; the core idea is reasonable but the comparison table has errors and protocol-mixing problems.","tokens_in":18212,"tokens_out":2604,"would_cite":false,"duration_ms":25094,"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 hybrid attention network beats prior bests on five X-ray landmark datasets.","keywords":["Anatomical Landmark Detection","Hybrid CNN-Transformer","Bi-Level Routing Attention","BiFormer","Attention Residual Module","Feature Fusion Correction Module","Dynamic Sparse Attention","X-ray image analysis"],"falsifier":"Retrain HTC, CeLDA, FARNet, and HYATT-Net from scratch with identical preprocessing, augmentation, and splits on ISBI2015 and CephAdoAdu using the released code and data, and compare MRE; if any baseline matches or beats the reported HYATT-Net numbers, the claimed state-of-the-art margins are not established.","tokens_in":17109,"feed_emoji":"🎯","tokens_out":14136,"duration_ms":113509,"temperature":0.7,"pith_summary":"The paper claims that a hybrid CNN-Transformer network, HYATT-Net, achieves state-of-the-art accuracy in anatomical landmark detection on X-ray images while remaining efficient on high-resolution inputs. Its core idea is to make attention query-aware and sparse: a bi-level routing mechanism first selects the most relevant image regions for each query and then applies token-level attention only inside those regions, which keeps global context modeling affordable. Combined with attention-augmented residual blocks and a feature fusion correction module, the network reports the lowest mean radial error on five diverse benchmarks, including 1.13 mm on ISBI2015 and 0.98 mm on the CephAdoAdu dataset. A sympathetic reader would take this as evidence that dynamic sparse attention, rather than ever-larger transformers, is a productive direction for medical landmark localization.","feed_headline":"Hybrid network beats prior bests on five X-ray landmark datasets","feed_subtitle":"Bi-level routing attention keeps global context modeling affordable, cutting mean radial error to 1.13 mm on ISBI2015.","key_machinery":"The carrying object is the BiFormer module built on Bi-Level Routing Attention (BRA), a two-stage sparse attention. BRA first computes patch-level queries and keys, forms a patch adjacency matrix, and applies a row-wise top-$k$ operator to select the $k$ most relevant patches per region; it then gathers only those key-value pairs and computes token-to-token attention with a depth-wise convolution (kernel size 5) as Local Context Enhancement. This mechanism is what makes global context modeling tractable at high resolution, and it is paired with two supporting modules: the Attention Residual Module, which wraps a dilated-convolution residual block with CBAM-style channel and spatial attention, and the Feature Fusion Correction Module, which concatenates global-pooled image features with the upsampled feature map to recover resolution. The deep-supervision scheme (MSE losses on three heatmap scales) ties the multi-scale predictions together.","core_discovery":"On the paper's own terms, the discovery is that query-aware dynamic sparsity in attention is sufficient to push anatomical landmark detection past prior state-of-the-art methods. HYATT-Net embeds the BiFormer module, which uses Bi-Level Routing Attention: the feature map is partitioned into patches, a coarse patch-level adjacency matrix is built from average queries and keys, and a row-wise top-$k$ operator keeps only the $k$ most relevant patches per query; token-to-token attention then runs on the gathered key-value pairs, supplemented by a depth-wise convolution for local context. This global feature is fed into an Attention Residual Module whose channel and spatial attention (CBAM) refine local features, and a Feature Fusion Correction Module pools global image context to restore resolution lost in downsampling. Training uses deep supervision with a mean-square error loss on heatmaps at three resolutions ($\\sigma=2,2,4$; weights $1,3,3$). The reported outcome is a mean radial error of 1.13 mm on ISBI2015 (HTC: 1.19 mm), 1.05 mm on ISBI2023, 0.98 mm on CephAdoAdu (CeLDA: 1.05 mm), 0.53 mm on the hand dataset, and 6.64 px on the pelvic dataset, with success detection rates higher than or close to the best baselines at most thresholds.","pith_inferences":["If the gains replicate under a common training protocol, the same recipe could transfer to other sparse-target dense prediction tasks such as fetal ultrasound landmark detection or surgical tool localization, where global geometry matters and images are high resolution; this goes beyond the paper's experiments.","A natural extension would be to condition the top-$k$ routing on an anatomical prior or a coarse landmark proposal map, which could further cut computation while holding accuracy; the paper does not test this.","Because the method helps most on the smallest training set (ISBI2015), a plausible but untested corollary is that dynamic attention's benefit grows as data become scarcer."],"forward_implications":["On the five evaluated X-ray benchmarks, HYATT-Net sets a new best mean radial error, so the method transfers across cephalometric, hand, and pelvic imaging with different resolutions and annotation counts.","Because the routing attention concentrates computation on the most relevant regions, high-resolution inputs (1024 by 1216) become affordable, which implies the approach can scale to other high-detail medical images without aggressive downsampling.","Ablations replacing BiFormer with a Swin or basic Transformer backbone lose over 1% SDR and more than 4% relative MRE, so the bi-level routing attention itself, not merely the transformer structure, drives the gain.","Adding CBAM and the feature fusion correction module each produces a small but consistent MRE reduction on ISBI2015, supporting the claim that global-guided local refinement and multi-scale fusion are complementary.","The released code and data allow the reported numbers to be reproduced and reused as a new baseline for later landmark detection work."],"supporting_citations":[{"why":"Supplies the Bi-Level Routing Attention (BRA) mechanism that is the core of the BiFormer module.","marker":"[41]"},{"why":"The hybrid Transformer-CNN baseline (HTC) whose reported MRE the paper beats on ISBI2015, ISBI2023, Hand, and Pelvic.","marker":"[24]"},{"why":"The prototypical network (CeLDA) that previously set state-of-the-art on the CephAdoAdu dataset and also provides that dataset's age-group protocol.","marker":"[46]"},{"why":"Provides the CBAM channel-and-spatial attention module used inside the Attention Residual Module.","marker":"[43]"},{"why":"The multi-scale feature aggregation/refinement baseline (FARNet) compared across head, hand, and pelvic datasets.","marker":"[16]"},{"why":"Defines the ISBI 2015 cephalometric benchmark dataset, its 19 landmarks, and the test protocol used in Table 1.","marker":"[44]"},{"why":"Defines the Cepha29/ISBI 2023 challenge dataset with 29 landmarks used to test cross-device generalization.","marker":"[45]"},{"why":"Supplies the pelvic X-ray landmark detection setup whose 512 by 512 resizing and data selection the paper follows.","marker":"[48]"}],"fun_headline_variants":["Hybrid attention net cuts landmark error to 1.13 mm on ISBI2015","BiFormer-based network beats prior bests on five X-ray datasets","HYATT-Net: efficient global attention for precise landmark detection","Hybrid CNN-Transformer sets new state of the art in landmark detection","Routing attention makes landmark detection faster and more accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed margins of improvement assume that the previously published baseline numbers cited with asterisks in Tables 1-3 were obtained under the same image resizing, augmentation, landmark sets, and train/test splits as the authors' own runs.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid attention net cuts landmark error to 1.13 mm on ISBI2015","BiFormer-based network beats prior bests on five X-ray datasets","HYATT-Net: efficient global attention for precise landmark detection","Hybrid CNN-Transformer sets new state of the art in landmark detection","Routing attention makes landmark detection faster and more accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1379,"prompt_tokens":1095,"completion_tokens":284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":192}},"tokens_in":711,"tokens_out":284,"duration_ms":3605,"temperature":1.0,"reasoning_tokens":192,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:35:14.808773+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain HTC, CeLDA, FARNet, and HYATT-Net from scratch with identical preprocessing, augmentation, and splits on ISBI2015 and CephAdoAdu using the released code and data, and compare MRE; if any baseline matches or beats the reported HYATT-Net numbers, the claimed state-of-the-art margins are not established.","supporting_citations":[{"cited_title":"Biformer: Vision transformer with bi-level routing attention","cited_arxiv_id":null,"evidence_quote":"Supplies the Bi-Level Routing Attention (BRA) mechanism that is the core of the BiFormer module."},{"cited_title":"In International Conference on Medical X","cited_arxiv_id":null,"evidence_quote":"The hybrid Transformer-CNN baseline (HTC) whose reported MRE the paper beats on ISBI2015, ISBI2023, Hand, and Pelvic."},{"cited_title":"In International Conference on Medical Image ComputingandComputer-AssistedIntervention ,pages 155–165","cited_arxiv_id":null,"evidence_quote":"The prototypical network (CeLDA) that previously set state-of-the-art on the CephAdoAdu dataset and also provides that dataset's age-group protocol."},{"cited_title":"Cbam: Convolutional block attention module","cited_arxiv_id":null,"evidence_quote":"Provides the CBAM channel-and-spatial attention module used inside the Attention Residual Module."},{"cited_title":"Feature aggregation and refinement network for 2d anatomical landmark detection","cited_arxiv_id":null,"evidence_quote":"The multi-scale feature aggregation/refinement baseline (FARNet) compared across head, hand, and pelvic datasets."},{"cited_title":"A benchmark for comparison of dentalradiographyanalysisalgorithms","cited_arxiv_id":null,"evidence_quote":"Defines the ISBI 2015 cephalometric benchmark dataset, its 19 landmarks, and the test protocol used in Table 1."},{"cited_title":"Pele scores: pelvic x-ray landmark detection with pelvis extraction and enhancement","cited_arxiv_id":null,"evidence_quote":"Supplies the pelvic X-ray landmark detection setup whose 512 by 512 resizing and data selection the paper follows."}],"review_version":1}