{"id":"2d6462a4-4ad9-4bf7-b528-1170b126b3f2","arxiv_id":"2502.09482","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An image-processing pipeline automatically extracts the convex ultrasound plane, fits it to an annulus sector, and linearises it for standardisation and augmentation.","lead":"This paper proposes an automatic method to find the fan-shaped ultrasound region in convex ultrasound images, represent it as a geometric sector, and transform it into a flat rectangle-like image. The goal is to make ultrasound images from different machines comparable and easier to use for AI training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The radial-boundary symmetry assumption in §2.3 (Eqs. 18–19) is load-bearing: when one side of the sector is shadowed or cropped, the method mirrors the other side, and the paper never quantifies the resulting parameter error.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing point: Section 2.3 assumes the centre of mass m splits the US plane symmetrically, so the better RANSAC fit is re-projected across the axis. My reading of the full text confirms that this is not a peripheral implementation detail. The annulus-sector parameters used for linearisation, origin O, opening angle θ, and inner/outer radii, are all computed from the two radial boundaries (Section 2.4). If one boundary is estimated by reflecting the other across an incorrect axis, the resulting error propagates into every scan line and into the invertibility claim. The paper's own public-data selection criteria exclude images too challenging for manual annotation and only include non-rotated or slightly rotated planes, so the reported MSE/MAAD numbers cannot establish robustness for the asymmetric cases the method is supposed to handle. The concrete test I propose would generate controlled asymmetric variants from the 30 private images, where ground truth is available, and measure how parameter error grows with asymmetry. If the test shows sensitivity, the paper should either restrict its claims to symmetric or minimally asymmetric planes or add an asymmetry detection step. This does not invalidate the method for the common symmetric case, but it means the central claim, as stated, is not yet supported. I therefore keep the reader's CONDITIONAL verdict unchanged. I do not see a more fundamental flaw: the pipeline is internally consistent, the RANSAC and connected-component steps are standard, and the reported experiments, while limited, are plausibly executed. The main overstatement is the phrase 'free of deformation,' which the authors themselves qualify by reporting a slight horizontal stretch and circularity of approximately 0.97; this is a wording issue rather than a structural one. Overall, the paper presents a promising approach whose main unresolved risk is the unquantified symmetry assumption, and a targeted sensitivity analysis would materially change the strength of the central claim.","tokens_in":12834,"tokens_out":3890,"duration_ms":41490,"concrete_test":"Take the 30 private intraoperative images and generate synthetic asymmetric variants of each: (a) horizontally translate the sector off-centre by ±5% and ±10% of image width, (b) occlude a wedge on one radial edge to simulate shadowing, and (c) crop one radial boundary. Run the proposed pipeline on these variants and compare estimated O, θ, rinner, and router against the ground-truth geometry of the transformed original. If key-point MSE or MAAD in θ grows monotonically with asymmetry magnitude beyond a few pixels or degrees, or if the reflected RANSAC line falls outside the masked plane, the symmetry assumption in Eqs. 18–19 is violated and the robustness claim requires explicit qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the method automatically extracts convex US planes from nonstandardised images, including corrupted and cropped cases, and linearises them with minimal deformation. The load-bearing step is the symmetry correction at the end of §2.3: after RANSAC-fitting both radial edges, the algorithm selects the fit with smaller residual and reflects it across the vertical line through m to obtain the other boundary (Eqs. 18–19). This is valid only when the annulus sector is symmetric about that line. The paper states this assumption explicitly but never quantifies its failure mode. Acoustic shadowing, off-centre probe contact, or lateral cropping break the symmetry; the special case in §3.2.2 handles only top-cropped images, not lateral asymmetry or rotation. Since O, θ, rinner, and router are all derived from the two radial boundaries, a biased boundary directly corrupts every downstream linearisation and augmentation. The public evaluation is partly shielded from this because images that were 'too challenging for human annotation' were excluded and only non-rotated or slightly rotated planes were annotated. Thus the evaluation cannot expose the failure that the method's own assumption acknowledges. The deformation experiment also reports a slight horizontal stretch after linearisation (circularity ≈0.97), qualifying the 'free of deformation' phrase in the conclusion, but the symmetry issue is more fundamental to correctness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a non-learning pipeline for standardising convex ultrasound images. The pipeline detects the ultrasound plane by histogram spike analysis and largest connected component, estimates annulus sector parameters (origin, opening angle, inner and outer radii) using a centre-of-mass line, RANSAC fits of the radial boundaries, and a symmetry-based re-projection of the better-fit line, and then extracts scan lines to linearise the plane via spline interpolation. Validation is performed on 30 intraoperative images with small reported key-point MSE and 0.30° mean absolute angular error, on 50 selected images from three public datasets, and through deformation, inversion, and radio-frequency experiments.","tokens_in":13133,"tokens_out":4898,"duration_ms":44773,"significance":"The idea of automatically recovering the convex US sector geometry without anatomy-specific priors or deep learning is practically appealing and, to my knowledge, not covered by prior work. If the robustness claims held, the method could reduce manual annotation effort and enable standardised augmentation and scan-line processing. The method is transparent, and the deformation, inversion, and RF experiments are useful first steps. However, the evaluation design leaves the central robustness claim insufficiently supported: the public dataset is selected to exclude challenging and rotated cases, and the most safety-critical assumption (radial-boundary symmetry) is never stress-tested. The contribution is promising but requires further validation before the paper's claims can be accepted.","major_comments":[{"comment":"The symmetry-based re-projection of the better RANSAC radial line is load-bearing for all downstream parameters: O (Eq. 21), θ, r_inner, and r_outer (Eq. 22) are computed from the two radial boundaries. The paper states the symmetry assumption but provides no sensitivity analysis for its violation. Acoustic shadowing, off-centre probe contact, or lateral cropping break the symmetry; the special crop handling in §3.2.2 only addresses top cropping, and the public evaluation explicitly excludes rotated and hard-to-annotate images. I request a quantitative failure-mode study, e.g. simulated shadowing/cropping on the controlled 30-image set with ground-truth parameters, reporting parameter errors as a function of asymmetry magnitude. Without this, the robustness claim in the abstract and Section 4 is not supported.","section":"§2.3, Eqs. (18)–(19)"},{"comment":"The public-data evaluation uses a selected subset of 50 images with two exclusion criteria: images too challenging for human annotation and rotated planes. These criteria directly remove the cases most likely to break the largest-connected-component and symmetry assumptions. The claim that the method is robust to nonstandardised images is therefore evaluated only on favourable cases. Please report the total number of images considered, the number excluded for each reason, and either results on a random or inclusive sample or an explicit statement that the method is intended only for non-rotated, human-annotatable images, with the claims revised accordingly.","section":"§3.2.2"},{"comment":"The inversion MSE of 0.0064 is a self-consistency check: the forward and inverse warps use the same estimated annulus parameters and the same interpolation scheme, so a low MSE only shows that the numerical transform is nearly invertible, not that the extracted geometry is accurate or that the augmentation preserves information. To support the conclusion that 'the image augmentation process does not affect the underlying US information', compare against an independent reconstruction, e.g. warp the linearised image back using ground-truth parameters, or report the MSE contribution attributable to parameter error separately. The circularity of 0.97 with visibly horizontal stretch in §3.3.1 should also be explicitly reconciled with the claim of deformation-free linearisation in the conclusion.","section":"§3.3.2"},{"comment":"The pipeline contains several manually chosen thresholds and assumptions: the z-score spike threshold factor, the half-height search range, the moving-average kernel, and the largest-connected-component rule. These choices precede the RANSAC and symmetry steps, so their impact on the final parameter estimates should be quantified. I ask for a sensitivity analysis on the controlled 30-image set, varying at least the spike threshold factor and the moving-average kernel over reasonable ranges and reporting the resulting key-point MSE. Without such an analysis, the reproducibility and robustness of the method are difficult to assess.","section":"§2.1–§2.2"}],"minor_comments":[{"comment":"The caption contains a typo: 'privaye' should be 'private'.","section":"Fig. 5 caption"},{"comment":"The trapezoidal-centre-of-mass formula is missing parentheses in the numerator; it should be written as Σ(iS[i]+(i+1)S[i+1])/2 over Σ(S[i]+S[i+1])/2.","section":"Eq. (9)"},{"comment":"The symbol S is used both for the accumulated row-difference vector and for the set of vertical intersections; renaming one of them would remove ambiguity.","section":"Eqs. (8) and (12)"},{"comment":"The expressions min(edgesl,i) and max(edgesl,i) are undefined; state that the minimum/maximum is taken over the image column coordinate of the edge set.","section":"Eq. (17)"},{"comment":"References [5] and [6] are duplicates of the same paper; one should be removed.","section":"References"},{"comment":"Several reported standard deviations exceed the mean (e.g. POCUS O: 23.73 ± 43.38); the authors should discuss the distribution and indicate whether a few outliers drive the result.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a useful gap and the geometric formulation is sound as far as it goes, but the evaluation is too favourably selected to support the general robustness claims. The symmetry-assumption failure mode and the self-referential inversion test are the two points that most need additional experiments. I would also encourage the authors to release code and exact parameter settings, since the current algorithmic description has enough ambiguities that independent reproduction would be difficult."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThe one thing to know: this is a genuinely useful, well-scoped paper on automatic convex ultrasound plane extraction, but its evaluation is narrower than its claims and the symmetry assumption in §2.3 is a real unquantified failure mode. It deserves serious review, not a desk reject.\n\nWhat's new is the task formulation itself. Automatically extracting the convex US plane as an annulus sector (origin, opening angle, inner/outer radii) and using that for scan-line extraction and linearisation is, as far as I know, new. The components are standard (connected components, RANSAC, splines), but the integration is sensible and the downstream applications are concrete. The private key-point results are good: MSE around 3–6 px, angular error 0.30°. The RF validation, comparing against a true linear image from RF data, is a nice independent check, even if it is one phantom slice and MS-SSIM of 0.69 is moderate.\n\nThe soft spots are in the evaluation and one load-bearing assumption. The public data selection explicitly excludes images too challenging for human annotation and excludes rotated planes. That means the robustness claims don't cover the hardest cases the paper says it handles. More importantly, the radial boundary refinement in §2.3 mirrors the better RANSAC fit across the centre of mass if the other side has higher residual. This assumes the sector is symmetric about m. Acoustic shadowing, off-centre contact, or lateral cropping break that symmetry, and since O, θ, rinner, and router are all derived from those boundaries, the error propagates everywhere. The paper states the assumption but never quantifies how bad the failure is. That is the main revision I'd ask for: a sensitivity analysis or at least a few asymmetric examples. The \"free of deformation\" phrase in the conclusion is also overclaimed; their own circularity result (0.97) shows slight horizontal stretch, and the inversion MSE is a self-consistency check, not accuracy.\n\nMinor: thresholds like the vertical search range and moving average kernel are arbitrary, and no code or data is released. The citation pattern looks fine.\n\nWho it's for: anyone working on ultrasound dataset curation, data augmentation, or scan-line processing. It deserves a serious referee; I'd recommend acceptance conditional on addressing the symmetry sensitivity and softening the claims. I'd cite it if I worked in this area.","headline":"A useful, well-scoped contribution to ultrasound standardisation, but the evaluation is too narrow and the symmetry assumption needs sensitivity analysis.","tokens_in":13608,"tokens_out":2178,"would_cite":true,"duration_ms":19167,"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":"The fan-shaped convex ultrasound plane can be automatically recovered as an annulus sector—origin, opening angle, inner and outer radii—and those parameters make linearisation, scan-line extraction, and invertible augmentation possible.","keywords":["ultrasound","annulus sector","convex probe","linearisation","scan line extraction","data augmentation","RANSAC","medical image standardisation"],"falsifier":"Take a set of convex ultrasound images whose planes are deliberately asymmetric—one edge obscured by acoustic shadowing or one side cropped—run the pipeline, and compare the estimated origin, opening angle, and radii against manual annotation; if the symmetry assumption is critical, parameter error should grow with the degree of asymmetry, and a synthetic annulus sector with a known partial occlusion could quantify the failure curve.","tokens_in":12660,"feed_emoji":"🩺","tokens_out":11302,"duration_ms":98768,"temperature":0.7,"pith_summary":"Ultrasound images from convex probes store the signal in a fan-shaped region whose curvature, position, and depth scale change with the machine, settings, and any cropping applied before publication. This paper argues that the fan can be represented by an annulus sector—a centre, an opening angle, an inner radius, and an outer radius—and that these four parameters can be recovered automatically from a still image using background-intensity detection, connected-component masking, a sliding-window centre estimate, and robust line fitting. If the recovery is accurate, the same parameters enable three practical operations without manual labelling: extracting the scan lines, straightening the fan into a rectangular linearised image, and applying invertible geometric augmentation. The authors validate the parameter estimates against manual annotations on private intraoperative brain ultrasound and on three public ultrasound sets, and show that linearisation keeps a circular object circular, round-trips back to the original image with negligible error, and agrees with a radio-frequency-derived ground truth. The broader point is that a purely geometric route can standardise disorganised and sparsely labelled convex ultrasound data before any learning-based processing begins.","feed_headline":"Automatic fan geometry standardises curved ultrasound images","feed_subtitle":"No manual labelling: the sector's origin, angle, and radii straighten the image, recover scan lines, and invert augmentation.","key_machinery":"The central object is the annulus sector, a fan parameterised by origin $O$, opening angle $\\theta$, inner radius $r_{inner}$, and outer radius $r_{outer}$; it reduces \"where is the ultrasound plane?\" to four numbers. The extraction pipeline is: binarise by detecting the background intensity spike in the histogram, isolate the largest connected foreground component as the plane, locate the inner arc and a central axis with a sliding-window row-difference accumulator and centre-of-mass estimate, fit the two radial edges with RANSAC, and reflect the better fit across the symmetry axis to repair the noisier side. The origin follows from the intersection of the two fitted lines via Cramer's rule, and the radii come from the distances from that origin to the inner and outer arc points on the central axis. Once the sector is known, scan lines are rays from the origin sampled by spline interpolation between $r_{inner}$ and $r_{outer}$, and the ratio of sector height to width sets the sampling density so the linearised image keeps a consistent resolution.","core_discovery":"The core discovery, stated on the paper's own terms, is that the visible convex ultrasound plane is a well-defined geometric object: all scan lines share a common origin and sweep through an opening angle between an inner and an outer arc. The paper shows that this annulus sector can be extracted automatically from ordinary B-mode images, even with superimposed GUI text, cropping, and acoustic artefacts, by masking the foreground, finding a central axis from the centre of mass, fitting the two radial legs with robust regression, and reflecting the better fit across the axis to compensate for the noisier side. From the estimated origin, angle, and radii, scan lines are generated by interpolating along radial rays and the fan is linearised into a rectangle whose columns follow the scan-line direction. The validation evidence is that key-point errors on private data are at the level of a few pixels with a mean angular error of 0.30 degrees, that noisy public data still yields angle errors near or below 3 degrees, that circularity of a ping-pong ball is 0.97 after linearisation, that the convex-to-linear-to-convex round trip has a mean squared error of 0.0064 pixels, and that multi-scale structural similarity to an RF-derived linear ground truth is 0.69. The paper's conclusion is that the method estimates the annulus sector parameters accurately enough to remove the need for manual annotation of the US sector in downstream standardisation.","pith_inferences":["A natural extension the paper does not pursue is using the estimated sector parameters as dataset metadata, so that every image in a benchmark carries its own origin, angle, and radii; this would make cross-dataset normalisation a lookup rather than a per-image fitting problem.","Because the linearisation is invertible with small round-trip error, the same geometry could support augmentations that stay on the ultrasound manifold, for instance adding realistic shadowing or depth-dependent attenuation before re-projecting to convex coordinates.","The method's sensitivity to the symmetry assumption could be tested on synthetic annulus sectors with controlled occlusions; such a study would show how much of the public-data error is explained by asymmetry rather than by resolution or annotation variance.","The representation is generic enough that it might transfer to other fan-shaped image geometries, such as phased-array or endoscopic ultrasound; the paper does not assert this transfer."],"forward_implications":["Convex ultrasound images from different machines can be normalised to a common fan geometry without manual annotation of the sector.","Scan lines can be extracted automatically, enabling explicit processing along the ultrasound beam direction in methods that previously required manual plane labelling.","The estimated annulus sector yields a linearised image whose column direction follows the scan lines, making rectangular-image operations like convolution or matrix multiplication applicable to convex data.","The linearisation is invertible: re-projecting the linear image back to convex coordinates reproduces the original image with a reported mean squared error of 0.0064 pixels, so the augmentation does not discard information.","The method is anatomy-agnostic and applies to non-rotated or slightly rotated convex and microconvex B-mode images, the typical clinical presentation; the authors note that reconstructed 3D ultrasound slices can fall outside this assumption."],"supporting_citations":[{"why":"Supplies the RANSAC robust line-fitting method used to estimate the two radial boundaries of the annulus sector.","marker":"[9]"},{"why":"Supplies the two-pass 4-connectivity connected-component labelling used to isolate the US plane from GUI elements and background.","marker":"[12]"},{"why":"Supplies one of the public ultrasound datasets used to evaluate parameter accuracy on noisy images.","marker":"[5]"},{"why":"Supplies a second public ultrasound dataset from a different acquisition setting for cross-machine evaluation.","marker":"[13]"},{"why":"Supplies a third public ultrasound benchmark used to test the method on cropped and slightly rotated planes.","marker":"[19]"},{"why":"Supplies the Procrustes analysis used to measure shape deformation of a circular object after linearisation.","marker":"[11]"},{"why":"Supplies the multi-scale structural similarity metric used to compare the linearised B-mode output with an RF-derived ground truth.","marker":"[27]"},{"why":"Supplies the reverberation example used to verify that linearisation projects artefacts vertically along scan lines.","marker":"[6]"}],"fun_headline_variants":["Curved ultrasound straightened automatically, no manual labels","Geometric annulus extraction standardises convex ultrasound fans","From fan to rectangle: auto geometry for ultrasound standardisation","Ultrasound sector geometry auto-detected for linearisation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the premise that the centre of mass of the masked ultrasound plane lies on a symmetry axis, so the better-fitting of the two radial edge lines can be mirrored to stand in for the other; if the visible plane is asymmetric because of shadowing, off-centre probe contact, or cropping, the estimated annulus sector inherits that error.","fun_headline_variants_meta":{"raw":{"variants":["Curved ultrasound straightened automatically, no manual labels","Geometric annulus extraction standardises convex ultrasound fans","From fan to rectangle: auto geometry for ultrasound standardisation","Ultrasound sector geometry auto-detected for linearisation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1533,"prompt_tokens":1026,"completion_tokens":507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":642,"tokens_out":507,"duration_ms":5436,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:19:01.078137+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of convex ultrasound images whose planes are deliberately asymmetric—one edge obscured by acoustic shadowing or one side cropped—run the pipeline, and compare the estimated origin, opening angle, and radii against manual annotation; if the symmetry assumption is critical, parameter error should grow with the degree of asymmetry, and a synthetic annulus sector with a known partial occlusion could quantify the failure curve.","supporting_citations":[{"cited_title":"Fischler and Robert C","cited_arxiv_id":null,"evidence_quote":"Supplies the RANSAC robust line-fitting method used to estimate the two radial boundaries of the annulus sector."},{"cited_title":"The connected-component labeling problem: A review of state-of-the-art algo- rithms","cited_arxiv_id":null,"evidence_quote":"Supplies the two-pass 4-connectivity connected-component labelling used to isolate the US plane from GUI elements and background."},{"cited_title":"Trahey, Sina Farsiu, and Mark L","cited_arxiv_id":null,"evidence_quote":"Supplies a second public ultrasound dataset from a different acquisition setting for cross-machine evaluation."},{"cited_title":"L¨ ubke, Sven Rothl¨ ubbers, Oudom Somphone, St´ ephane Allaire, Muyinatu A","cited_arxiv_id":null,"evidence_quote":"Supplies a third public ultrasound benchmark used to test the method on cropped and slightly rotated planes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Procrustes analysis used to measure shape deformation of a circular object after linearisation."},{"cited_title":"Simoncelli, and Alan Conrad Bovik","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-scale structural similarity metric used to compare the linearised B-mode output with an RF-derived ground truth."},{"cited_title":"Accelerating detec- tion of lung pathologies with explainable ultrasound image analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the reverberation example used to verify that linearisation projects artefacts vertically along scan lines."}],"review_version":1}