{"id":"8aa3365b-0a82-4d99-9166-9a22d330c180","arxiv_id":"2506.23721","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An AR-assisted ultrasound pipeline for real-time kidney segmentation and automated volume measurement is built and tested with four deep learning models.","lead":"This paper presents an augmented reality ultrasound system that streams kidney images to a deep learning segmentation model and displays the results on a HoloLens 2 headset. The system automatically estimates kidney volume and could help clinicians keep their eyes on the patient while scanning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Automated volume measurement is the central claim, but volume is never validated and the 'capsule' mask used for it is never defined; Table 2 shows systematic errors even with ground-truth segmentations.","rationale":"The reader's weakest_assumption identifies the same load-bearing gap: the capsule mask is undefined and the PCA bounding box is assumed to reproduce manual caliper measurements. My analysis sharpens this by noting that even the Ground Truth row in Table 2 shows systematic errors of about 4 mm in each linear dimension, which indicates the measurement algorithm itself is not a validated proxy for manual measurements. Additionally, the paper claims 'automated kidney volumetric measurements' but never reports a volume error; only length, width, and thickness are evaluated, and volume is computed multiplicatively so its error is not bounded by the linear errors. The in-house dataset of 20 images from 5 healthy volunteers is far too small to establish interobserver-level accuracy, and the paper's own Future Work admits that comprehensive validation is still required. These concerns are specific and addressable, consistent with the reader's CONDITIONAL verdict, so no verdict change is warranted. The segmentation comparison on the public OpenUS dataset is competently executed, and the inference-time measurements are transparent; the issue is not the segmentation quality but the unvalidated measurement pipeline built on top of it.","tokens_in":9266,"tokens_out":2418,"duration_ms":28772,"concrete_test":"Ask the authors to specify the exact capsule mask construction (e.g., union of cortex and central complex) and to release the code path implementing Eqs. 2-4. Then independently recompute Table 2's Ground Truth row on the same 20 physician-annotated in-house images, and additionally compute the ellipsoid volume error against a reference (e.g., CT volumetry or a calibrated phantom). If the recomputed length/width/thickness errors exceed the reported values, or if the volume error exceeds clinically acceptable interobserver variability, the central volumetric-accuracy claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is that the pipeline delivers automated kidney volumetric measurements with errors close to clinical interobserver variability. For that claim to hold, three things must be true: (1) a 'capsule' mask can be reliably formed from the two segmented classes (cortex and central complex); (2) the PCA-aligned bounding rectangle in Eq. 2-4 yields the same length, width, and thickness that a sonographer records with manual calipers; and (3) the resulting ellipsoid volume (Eq. 1) is accurate enough for clinical use. The weakest link is that none of these is actually established. The capsule construction is described only as 'the capsule segmentation is used to extract kidney dimensions' (General framework) with no definition or formula linking it to the cortex/central complex outputs. Even if that gap were filled, Table 2's 'Ground Truth' row shows that physician-annotated segmentations run through the automatic algorithm produce length error 4.22±1.52 mm, width error 3.84±2.36 mm, and thickness error 4.16±1.39 mm against device measurements; the paper attributes this to misalignment between manual calipers and post-processed segmentation, but this is precisely the comparison that must be validated, not assumed. No volume error is reported anywhere, even though volume is the headline output. The in-house test set is only 20 images from 5 young, low-BMI volunteers, so the 'close to interobserver variability' claim rests on a very small, demographically narrow sample. The paper itself concedes in Future Work that comprehensive end-user clinical validation is still required. These are addressable but currently unresolved gaps, so the conditional verdict is appropriate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript integrates deep learning based semantic segmentation with augmented reality on HoloLens-2 for kidney ultrasound, proposing two streaming pipelines (wireless API-based and wired video grabber) and an automatic measurement workflow that computes length, width, thickness, and ellipsoid volume. Four segmentation models (nnU-Net, Segmenter, YOLO v11 with MedSAM, and YOLO v11 with LiteMedSAM) are compared on the public OpenUS Kidney Dataset with 5-fold cross-validation, and measurement errors are reported on a 20-image in-house dataset from five volunteers. The authors claim that Segmenter enables real-time inference, that the system automates kidney volumetric measurements, and that measurement errors are close to clinical interobserver variability, with the pipeline released as open source.","tokens_in":9571,"tokens_out":4399,"duration_ms":49658,"significance":"The paper has clear strengths: it uses a public dataset, open-source models, 5-fold cross-validation, and it releases an open-source pipeline, which supports reproducibility. If the measurement pipeline were properly validated, the work would be a useful practical contribution to point-of-care ultrasound and AR-assisted workflows. However, the central claims about automated volumetric measurements and real-time end-to-end operation are currently under-supported: the capsule segmentation used for measurements is never defined, volume accuracy is never evaluated, interobserver variability is not measured on the same protocol, and only model inference time is reported rather than end-to-end latency.","major_comments":[{"comment":"The 'capsule segmentation' is load-bearing but never defined. The models output two classes, renal cortex and central complex, and the General framework states only that 'The capsule segmentation is used to extract kidney dimensions.' No rule is given for deriving a capsule mask from these two classes, yet Eqs. (2)-(4) are applied to a binary matrix whose provenance is unspecified. Please specify the capsule construction, for example as a union of the two classes, a contour, or a morphologically closed mask, and justify that it agrees with the kidney boundary used by sonographers.","section":"General framework; DL-Based Kidney Segmentation"},{"comment":"The Ground Truth row of Table 2 shows that the automatic algorithm applied to physician-annotated segmentations already has mean errors of 4.22 mm for length, 3.84 mm for width, and 4.16 mm for thickness relative to device calipers. The paper attributes this to misalignment between manual point-of-care measurements and post-processed segmentation, but this is precisely the comparison that must be validated, not assumed. Without an interobserver variability measurement on the same in-house protocol, or clearly matched literature values for the same measurement axes, the statement that errors are comparable to interobserver variability is not supported. The standard deviations in the model rows are also large, up to 17 mm, so reporting means alone overstates consistency.","section":"Table 2; Measurements evaluations"},{"comment":"Volume is the headline output, but no volume error is reported anywhere. Equation (1) combines length, width, and thickness from two views, yet the Measurements evaluations section gives only linear dimension errors. To support the abstract claim of automated kidney volumetric measurements, the authors should report volume errors against a reference, such as manual ellipsoid volume from the same images or a volumetric standard, including bias and limits of agreement. The assumption that the PCA-aligned bounding rectangle in Eqs. (2)-(4) recovers the clinician-defined kidney axes is also asserted rather than tested; the Ground Truth row of Table 2 suggests a systematic offset that should be investigated.","section":"Automatic Measurements; Eq. (1)"},{"comment":"The real-time claim is based on model inference time only. Table 1 shows that only Segmenter (23.4 ms) clearly meets a typical real-time threshold, while nnU-Net (338 ms) and YOLO with MedSAM (334.6 ms) do not, and YOLO with LiteMedSAM is borderline and variable (76.8 +/- 38.5 ms). More importantly, no end-to-end latency is reported for either pipeline, including acquisition, streaming, pre- and post-processing, and AR rendering. Without this measurement and without a stated real-time criterion, the manuscript's claim of real-time automated volumetric measurements is not established for the complete system.","section":"Table 1; Data Streaming Between US Device and PC"}],"minor_comments":[{"comment":"The sentence 'nnU-Net and Segmenter achieve higher accuracy for coronal-view images than for coronal images' should read 'than for transverse-view images', as written it is contradictory.","section":"Measurements evaluations"},{"comment":"The phrase 'volume volume [5]' contains a duplicated word and should be corrected to 'volume [5]' or 'volume (Eq. 1)'.","section":"General framework"},{"comment":"The sentence 'Points in P are rotated by -theta to align with the axes normal axes' is garbled; please rewrite as 'to align with the principal axes'.","section":"Automatic Measurements"},{"comment":"The text states that all models were trained and validated on 512 x 512 resolution images, but the native resolution of the OpenUS Kidney Dataset images is not described; please specify the resampling or padding procedure used.","section":"Implementation details"},{"comment":"The label 'Ground Truth' for the row that applies the automatic algorithm to physician-annotated segmentations is ambiguous; consider renaming it to 'Annotated masks + automatic algorithm' or 'Manual segmentations with automatic measurement' to distinguish it from the physician's device measurements.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an imaging-focused journal and the open-source pipeline is a useful contribution. The main risk is overclaiming clinical readiness: the volume claim and the interobserver-variability comparison need direct evidence, and the real-time claim needs end-to-end latency. These are fixable with additional experiments and a more cautious interpretation, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper is an honest integration effort, not a research breakthrough. It compares four open-source segmentation models on a public kidney ultrasound dataset and packages two AR streaming pipelines for the HoloLens 2. If you work on point-of-care ultrasound or AR-assisted imaging, it's worth a look for the open-source release and the clear inference-time table. The segmentation evaluation is the strongest part: 5-fold cross-validation, DICE/IoU/mAP, and the trade-off between nnU-Net's accuracy and Segmenter's real-time speed is clearly presented.\n\nThe second thing: the paper's central claim about automated volumetric measurements does not hold up as written. The 'capsule segmentation' that feeds the PCA-based length/width/thickness extraction is never defined. As far as I can tell from the methods, the models output cortex and central complex; there is no formula or morphological description showing how those two classes become a capsule mask. That is not a small omission because Table 2 shows that even physician-annotated segmentations run through the automatic algorithm produce length errors around 4.2 mm against device caliper measurements. The paper calls this misalignment, but since the manual measurement is the clinical ground truth, that 4 mm gap is exactly what the system must beat, not explain away. And volume – which is the actual output in the abstract and the ellipsoid formula – is never validated at all. There is no volume error, no comparison to CT or a reference standard.\n\nAlso soft, in order: the in-house test set is 20 images from five young, low-BMI volunteers, so the generalization claims are thin; end-to-end latency is not measured (only model inference time); and there is no user study, even though the ergonomics and 'switching focus' benefits are a large part of the motivation. The authors do acknowledge most of this in the Future Work section, which is why I don't read this as a cynical paper. They just overreach in the abstract and conclusion when they say error rates are 'close to clinical interobserver variability.'\n\nIf I were handling this, I'd send it to review with a clear mandate: define the capsule mask, validate volume against a real reference (CT or at least a stated standard), and either measure end-to-end latency or drop the real-time claim for the wired pipeline. The segmentation comparison and the open-source release justify the referee time. It is a conditional accept, not a reject.","headline":"Reasonable AR-US integration with a solid segmentation comparison, but the headline measurement claim rests on an undefined 'capsule' mask and no volume validation.","tokens_in":10133,"tokens_out":1804,"would_cite":false,"duration_ms":18377,"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":"The paper claims that pairing real-time deep-learning segmentation with an augmented-reality headset can automate kidney length, width, thickness, and volume measurements from ultrasound with errors of only a few millimeters.","keywords":["augmented reality","kidney ultrasound","semantic segmentation","real-time inference","volumetric measurements","PCA bounding box","ellipsoid volume","point-of-care ultrasound"],"falsifier":"Run the automatic bounding-box measurement on physician-annotated kidney masks across diverse shapes—normal adult, pediatric, polycystic, and shrunken kidneys—and compare directly with the same clinicians' manual caliper measurements; if length or width errors from perfect masks exceed the few-millimeter target in a substantial fraction of cases, the geometric proxy rather than the segmentation would be the failing link.","tokens_in":9046,"feed_emoji":"🩺","tokens_out":5755,"duration_ms":57965,"temperature":0.7,"pith_summary":"This paper sets out to show that ultrasound kidney measurements can be automated end-to-end by combining real-time deep-learning segmentation with an augmented-reality headset display, so that a sonographer sees the kidney mask, length, width, thickness, and estimated volume directly in their field of view while scanning. The authors report that the automated measurements land within a few millimeters of physician measurements, close to published clinical interobserver variability, despite training on a relatively small public dataset and testing on an in-house set outside the training distribution. The motivation is practical: manual volumetric kidney measurement is time-consuming and operator-dependent, and the constant gaze shifts between patient and screen add cognitive load. If the pipeline holds up in larger clinical validation, it would let point-of-care users concentrate on image interpretation instead of caliper placement.","feed_headline":"Deep learning plus AR measures kidney ultrasound to within millimeters","feed_subtitle":"Real-time overlays on a headset let sonographers see automated kidney measurements without looking away from the patient.","key_machinery":"The measurement machinery is a PCA-aligned bounding box computed from a binary 'capsule' kidney mask. Given the active pixels of the mask, principal component analysis gives the kidney's orientation; the mask is rotated to align with the axes, the minimum and maximum extents define the bounding rectangle, and rotating back yields the length, width, and thickness. These feed the ellipsoid volume formula $V = \\frac{\\pi}{6} L W T$. The same mask comes from two segmentation classes, cortex and central complex, which the evaluated models (nnU-Net, Segmenter, YOLO v11 with MedSAM or LiteMedSAM) predict per frame.","core_discovery":"The central claim is that a binary capsule segmentation of the kidney—formed from the cortex and central complex classes produced by the segmentation model—can be turned into clinically usable dimensions by aligning a bounding box to the segmentation's principal axis via PCA, computing length, width, and thickness from that box, and estimating volume with the ellipsoid formula $V = \\frac{\\pi}{6} L W T$. Feeding this pipeline with segmentation from nnU-Net, Segmenter, or a YOLO-plus-MedSAM cascade yields length, width, and thickness errors of a few millimeters relative to physician measurements on an in-house dataset, with Segmenter fast enough (about 23 ms per frame) for real-time use. The paper further claims that projecting this overlay onto an AR headset and providing two streaming pipelines (wireless for API-connected scanners and wired video capture for any ultrasound machine) makes the automation practical in point-of-care settings. The strongest version of the claim is that automated measurement error is close to clinical interobserver variability even with limited training data and on an out-of-distribution test set.","pith_inferences":["If the capsule mask is obtained deterministically from the cortex and central-complex outputs, then measurement error is bottlenecked by segmentation quality; improving transverse-view segmentation should cut width and thickness errors more than length errors.","The bounding-box proxy ties automated dimensions to the shape of the segmented mask, so kidneys with irregular outlines or poor acoustic windows may need an alternative measurement definition before the pipeline generalizes to pathological cases.","Because even physician-annotated masks run through the algorithm show roughly 4 mm length error, part of the reported error is the geometric proxy itself rather than the deep-learning models; aligning the measurement protocol with clinical caliper conventions could shrink the gap without retraining.","A direct user study comparing gaze behavior, task time, and diagnostic confidence with and without the AR overlay would test the paper's claim that the headset reduces cognitive load, which the current evaluation does not measure."],"forward_implications":["Sonographers could keep their eyes on the patient while live kidney dimensions and volume appear as an overlay, reducing gaze shifts and cognitive load.","Measurement error in the few-millimeter range would make automated calipers a viable assist for routine kidney volume assessment, with physicians correcting the overlay in seconds.","Because Segmenter runs at around 23 ms per frame, near-real-time feedback is achievable on current hardware; nnU-Net remains the accuracy reference when speed is less critical.","The wired video-capture pipeline extends the same AR assistance to any ultrasound machine with video output, not just scanners with a programming interface.","A public dataset and open models make the pipeline reproducible and directly extendable to training tools for novice sonographers."],"supporting_citations":[{"why":"Supplies the expert-annotated kidney ultrasound dataset used to train and evaluate all four segmentation models.","marker":"[23]"},{"why":"Provides the nnU-Net self-configuring segmentation framework that achieves the highest accuracy in the comparison.","marker":"[24]"},{"why":"Provides the Segmenter transformer model that delivers real-time inference at about 23 ms per frame.","marker":"[25]"},{"why":"Provides the YOLO v11 detector used in the two-stage cascade to localize the kidney before segmentation.","marker":"[26]"},{"why":"Provides MedSAM, the bounding-box-prompted segmentation model used in the cascade.","marker":"[27]"},{"why":"Provides LiteMedSAM, the lightweight cascade variant evaluated for near-real-time inference.","marker":"[28]"},{"why":"Supplies the ellipsoid formula used to estimate kidney volume from length, width, and thickness.","marker":"[30]"},{"why":"Documents sonographic renal length interobserver variation in children, anchoring the clinical variability the automated errors are compared against.","marker":"[7]"},{"why":"Reports interobserver variation in adult kidney size measurements, another baseline for the few-millimeter claim.","marker":"[9]"},{"why":"Measures reliability of ultrasound renal length in adults, supporting the comparison to clinical measurement variability.","marker":"[10]"}],"fun_headline_variants":["AR and deep learning measure kidney ultrasound in real time","Real-time kidney measurements via AR-assisted deep learning","AI plus AR gives instant kidney ultrasound measurements","Millimeter-accurate kidney ultrasound readings with AR and AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the length, width, and thickness a sonographer records with manual calipers can be reproduced by a PCA-aligned bounding box around a binary capsule mask; if clinical measurement conventions do not match what that bounding box measures, the automated numbers will be biased even when the segmentation is perfect.","fun_headline_variants_meta":{"raw":{"variants":["AR and deep learning measure kidney ultrasound in real time","Real-time kidney measurements via AR-assisted deep learning","AI plus AR gives instant kidney ultrasound measurements","Millimeter-accurate kidney ultrasound readings with AR and AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1519,"prompt_tokens":995,"completion_tokens":524,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":462}},"tokens_in":611,"tokens_out":524,"duration_ms":5559,"temperature":1.0,"reasoning_tokens":462,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:33:26.122014+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the automatic bounding-box measurement on physician-annotated kidney masks across diverse shapes—normal adult, pediatric, polycystic, and shrunken kidneys—and compare directly with the same clinicians' manual caliper measurements; if length or width errors from perfect masks exceed the few-millimeter target in a substantial fraction of cases, the geometric proxy rather than the segmentation would be the failing link.","supporting_citations":[{"cited_title":"The open kidney ultrasound data set,","cited_arxiv_id":null,"evidence_quote":"Supplies the expert-annotated kidney ultrasound dataset used to train and evaluate all four segmentation models."},{"cited_title":"Segmenter: Transformer for se- mantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the Segmenter transformer model that delivers real-time inference at about 23 ms per frame."},{"cited_title":"Rep-medsam: Towards real-time and universal medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides LiteMedSAM, the lightweight cascade variant evaluated for near-real-time inference."},{"cited_title":"A simple method to estimate renal volume from computed tomography,","cited_arxiv_id":null,"evidence_quote":"Supplies the ellipsoid formula used to estimate kidney volume from length, width, and thickness."},{"cited_title":"Interobserver and intraobserver variations in sonographic renal length measurements in children","cited_arxiv_id":null,"evidence_quote":"Documents sonographic renal length interobserver variation in children, anchoring the clinical variability the automated errors are compared against."},{"cited_title":"Intraobserver and interobserver variations in sonographic measurements of kidney size in adult volunteers: a comparison of linear measurements and volumetric estimates,","cited_arxiv_id":null,"evidence_quote":"Reports interobserver variation in adult kidney size measurements, another baseline for the few-millimeter claim."},{"cited_title":"How reliable are ultrasound measurements of renal length in adults?","cited_arxiv_id":null,"evidence_quote":"Measures reliability of ultrasound renal length in adults, supporting the comparison to clinical measurement variability."}],"review_version":1}