{"id":"6ed4bcca-ab4a-46c6-8384-033184edeefd","arxiv_id":"2411.11190","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DeepSPV is the first deep learning pipeline to estimate spleen volume from 2D ultrasound images, achieving 83.0% mean relative volume accuracy in a synthetic-image full-pipeline test and outperforming standard clinical formulas in that test.","lead":"DeepSPV estimates 3D spleen volume from one or two standard 2D ultrasound images using a U-Net segmenter and a variational autoencoder that regresses volume from the segmented shape. The best configuration reports 86.62% and 92.58% mean relative volume accuracy on single-view and dual-view inputs, but the complete pipeline is validated only on synthetic ultrasound images.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Full-pipeline MRVA of 83.0% is measured on synthetic US images that the U-Net segments much more accurately than real clinical US (Dice 0.975 vs 0.929), and the human-expert comparator is carried over from CT-based measurements, so the claimed clinical-level accuracy is not established.","rationale":"The reader's verdict is CONDITIONAL, and my stress test agrees. The single most load-bearing assumption is that the synthetic US images used for the full-pipeline evaluation are representative of real clinical US in the target population. The paper's own quantitative results undercut this assumption: the U-Net achieves Dice 0.9746/HD 4.09 mm on the synthetic test images used for the 83.0% MRVA, versus Dice 0.9285/HD 11.11 mm on real clinical US. The full-pipeline volume accuracy is measured on the easier synthetic distribution, so the 3.62-point drop from ground-truth segmentations to the full pipeline is likely an optimistic lower bound for the segmentation-induced error on real images. The age mismatch (adult CT training vs paediatric SCD target) compounds this, as the paper acknowledges. The comparison to human experts in Table 5 is also not an apples-to-apples measurement on the synthetic US test set; the 68.54% value comes from the CT-based experiment. None of this is internal inconsistency or fraud; the internal experiments on CT-derived segmentations are sound, and the paper honestly lists the lack of paired data as a limitation. But it means the clinical claim is conditional on future validation. Therefore UNCHANGED (still CONDITIONAL). The concrete test—paired real US with volumetric ground truth—would settle the question.","tokens_in":28023,"tokens_out":10017,"duration_ms":94155,"concrete_test":"Acquire paired data from N≥50 paediatric SCD patients with a clinically acquired coronal 2D US image and a CT/MRI volume within 30 days; compute spleen volume from the 3D scan as ground truth, run the full DeepSPV pipeline on the US image, and report MRVA and U-Net Dice on the real US images. If the observed MRVA is materially below 83.0% (or if the Dice matches the Table 4 real-US value of 0.9285 rather than the synthetic 0.9746, and the resulting segmentation-error penalty exceeds the 3.62 points seen on synthetic images), the reported full-pipeline accuracy is not representative of clinical performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DeepSPV reaches 83.0% MRVA on the full pipeline from a single 2D US image is supported only by synthetic images whose segmentation is substantially easier than real clinical US. Table 4 reports U-Net Dice/HD of 0.9746/4.09 mm on synthetic US generated from CT-derived layouts, versus 0.9285/11.11 mm on the real US test set. Table 5 then uses exactly those synthetic test images to measure the 83.0% MRVA and the 3.62-point drop relative to ground-truth CT segmentations. Because the real US test set is the target population (paediatric SCD patients) and the synthetic test set is derived from adult CT spleens, the easier segmentation on synthetic images likely understates the volume error that would occur on real images. The RVAE was trained on adult CT segmentations; the paper acknowledges the adult/pediatric mismatch and asserts no significant impact without evidence. Additionally, the 'Human Experts 68.54%' cited in Table 5 is the value from the CT-segmentation experiment (Table 2), not a measurement on the synthetic US test set, so the full-pipeline claim of surpassing human experts is an apples-to-oranges comparison. The 83.0% is therefore an optimistic upper bound, not a clinical performance estimate; real paired US/volumetric data are needed to establish the claimed accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepSPV, a two-stage deep learning pipeline that estimates 3D spleen volume from single or dual 2D ultrasound images: a U-Net segments the spleen in each 2D ultrasound image, and a VAE-based regressor (RVAE) maps the resulting segmentation(s) to a volume estimate. Because no paired real ultrasound volumes are available, the authors also introduce an ultrasound semantic diffusion model (USDM) that synthesizes realistic 2D ultrasound images from CT-derived spleen segmentations, and they evaluate the full pipeline on these synthetic images. On CT-derived 2D segmentations, the RVAE achieves 86.62% and 92.58% mean relative volume accuracy (MRVA) for single-view and dual-view settings, outperforming Pix2Vox++, ResNet/VGG/DenseNet regressors, and clinical linear-regression formulas. On synthetic ultrasound images, the full pipeline achieves 83.0% MRVA, which the authors compare with a human-expert MRVA of 68.54% and use to claim that the pipeline surpasses human performance.","tokens_in":28307,"tokens_out":4270,"duration_ms":40712,"significance":"If validated on real clinical data, the CT-based volume-estimation experiments would be a solid contribution: the 5-fold stratified evaluation, the inclusion of multiple baselines, the pairwise significance tests, and the confidence-interval and interpretability analyses are strengths. The public synthetic ultrasound dataset is also a useful resource for the community. However, the central clinical claim rests on the full-pipeline evaluation on synthetic ultrasound, and that evaluation has two serious gaps: the synthetic test images are segmented much more accurately than real ultrasound images by the same U-Net, and the human-expert comparator is not measured on the synthetic test set. The paper therefore does not yet establish the claimed clinical-level accuracy from 2D ultrasound.","major_comments":[{"comment":"The full-pipeline MRVA of 83.0% is computed on synthetic ultrasound images generated from the held-out CT spleen segmentations (Section 4.2.2, Table 1). The 'Human Experts' value of 68.54% in Table 5 is not a measurement on those synthetic images; it is the single-view human-expert MRVA from Table 2, obtained from manual linear-regression measurements on CT-derived slices. The comparison is therefore apples-to-oranges and does not support the claim that the full pipeline surpasses human experts on the same data. Please either re-measure human performance on the synthetic test images or explicitly label the Table 5 human-expert row as a CT-based reference and remove the 'surpasses human experts' claim for the full pipeline.","section":"§5.2.4, Table 5"},{"comment":"The segmentation U-Net achieves Dice 0.9746 and HD 4.09 mm on synthetic ultrasound images generated from synthetic layouts, but only 0.9285 and 11.11 mm on the real ultrasound test set. The full-pipeline volume error is measured on the synthetic test images, so the reported 83.0% MRVA is an optimistic upper bound for real clinical images. The paper should either quantify how the observed segmentation gap affects volume estimation (e.g., by running the volume estimator on real-US segmentations with an alternative volume source) or explicitly state that the true MRVA on real ultrasound may be lower. This issue directly affects the paper's central quantitative claim.","section":"§5.2.3, Table 4 vs §5.2.4"},{"comment":"The synthetic layouts are created by rotating a 2D CT spleen segmentation 90 degrees anti-clockwise and manually overlaying it on a real ultrasound cone, with no validation that this reproduces the clinically recorded coronal view, which is aimed at the longest spleen length. In addition, the coronal slice selected from each CT volume is the slice with the largest cross-sectional area, and it is assumed to correspond to the clinical coronal ultrasound view. If this assumption fails, the synthetic test set does not represent the clinical acquisition protocol, further undermining the transferability of the 83.0% MRVA. Please provide evidence for the anatomical correspondence or temper the claims about clinical workflow integration.","section":"§3.3 and §4.1.2"},{"comment":"The Discussion acknowledges that the CT data are from adults while the ultrasound data are from paediatric SCD patients, and then asserts 'we do not expect this difference to have a significant impact on our results and conclusions' without supporting evidence. Since the volume estimator is trained exclusively on adult CT segmentations and the target population is paediatric, this is an unsupported assumption that affects the generalisability claim. Please either provide quantitative evidence (e.g., age-stratified sensitivity analysis or literature-based volume ranges) or explicitly present the adult/pediatric mismatch as an unresolved limitation rather than a benign feature.","section":"Section 6, last paragraph"}],"minor_comments":[{"comment":"The abstract reports '86.62%/92.5%' but Table 2 reports 92.58%; please make these values consistent.","section":"Abstract and §5.2.1"},{"comment":"The reparameterisation equation uses the notation 'z=µ+ζ J σ' with an undefined symbol; replace it with a standard element-wise product, e.g., 'z = µ + ζ ⊙ σ'.","section":"§3.2.1"},{"comment":"The figure contains the typo 'valiation' instead of 'validation'.","section":"Fig. 1"},{"comment":"The Data Availability statement says the synthetic dataset 'will be made publicly available on paper publication', while the abstract states 'We also make our synthetic spleen ultrasound dataset publicly available'; please align the two statements.","section":"Data Availability"},{"comment":"The reference list contains duplicate entries for Altini et al. (2022) and Holmström et al. (2022); please deduplicate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to be of interest to the medical ultrasound and image-analysis community, and the CT-based volume estimation experiments are competently executed. The main concern is that the abstract and conclusion present the synthetic-image evaluation as establishing clinical-level accuracy, which the evidence does not support. Please ensure the revised version clearly distinguishes a proof-of-concept on synthetic data from a clinically validated performance claim, and either re-measure the human-expert comparator on the synthetic test set or remove the 'surpasses human experts' claim for the full pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's useful core is a well-run comparison of VAE-based volume estimation from 2D spleen segmentations, and that part mostly holds up. The full-pipeline clinical claim does not.\n\nWhat is genuinely new: first deep-learning pipeline for spleen volume from 2D ultrasound; the synthetic US dataset is public; the study compares against Pix2Vox++, three CNNs, and standard clinical formulas on a sensible 5-fold split. RVAE's 86.62%/92.58% MRVA on CT-derived segmentations beat the baselines, and the robustness-to-rotation tests are a nice touch. The interpretability analysis is basic but reasonable.\n\nWhere it gets soft. The full-pipeline 83.0% MRVA is measured on synthetic US images generated from the same CT segmentations that provide ground-truth volumes. Those synthetic images are substantially easier for the U-Net than real clinical US: Dice 0.9746 vs 0.9285, HD 4.09mm vs 11.11mm. So 83.0% is an optimistic upper bound, not a real-world error estimate. The 'Human Experts 68.54%' in Table 5 is carried over from the CT-segmentation experiment, not measured on the synthetic test set, so the abstract's claim of surpassing human experts on the full pipeline is apples-to-oranges. The authors also assume the CT slice with largest cross-sectional area matches the clinical coronal view of longest spleen length; no paired real US with volumetric ground truth anywhere tests that assumption. The adult CT / paediatric US mismatch is acknowledged but not examined. The confidence intervals are called 95% but MCIA comes in at 83-84%, so they are not well calibrated. The full pipeline is also only evaluated single-view; dual-view results stay on CT segmentations. None of this kills the method; it means the validation is incomplete and the language overclaims.\n\nWho it's for: researchers working on ultrasound-based volumetry or splenomegaly monitoring. The paper deserves a serious referee; the internal experiments are solid enough to merit full review. The authors should be asked to validate on paired real US/CT data, report segmentation-error propagation on real US, and rerun the human comparison on the same test data.","headline":"Useful first system for spleen volume from 2D US, but the clinical-level accuracy claims rest on synthetic images that are easier to segment than real US, so the headline numbers are optimistic.","tokens_in":28894,"tokens_out":3245,"would_cite":true,"duration_ms":31862,"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":"DeepSPV estimates spleen volume from one or two 2D ultrasound images, and the best model beats human expert accuracy on synthetic ultrasound.","keywords":["splenomegaly","2D ultrasound","deep learning","volume estimation","variational autoencoder","diffusion model","synthetic ultrasound","sickle cell disease"],"falsifier":"Collect paired 2D ultrasound images and CT or MRI spleen volumes from the same patients, run DeepSPV on the real ultrasound images, and compare mean relative volume accuracy with the CT volumes; if the accuracy falls substantially below the reported 83% or below the human-expert baseline, the synthetic-to-real transfer assumption fails.","tokens_in":27785,"feed_emoji":"🩺","tokens_out":7255,"duration_ms":60666,"temperature":0.7,"pith_summary":"Spleen volume is the gold-standard measure of spleen size, but it normally requires CT or MRI, which are often unavailable where sickle cell disease is common. This paper argues that a deep learning pipeline called DeepSPV can recover 3D spleen volume from one coronal 2D ultrasound image, or from coronal plus transverse views, using only the segmentations visible in standard clinical images. On CT-derived segmentations, the best model reaches 86.62% mean relative volume accuracy with a single view and 92.58% with two views, compared with 68.54% and 81.60% for human experts using standard linear-regression formulas. On a full pipeline test that first segments synthetic ultrasound images generated by a diffusion model, then estimates volume, it reaches 83.0% from one view. If these numbers transfer to real patients, spleen volume could become a practical biomarker in routine ultrasound settings.","feed_headline":"Deep learning reads spleen volume from routine 2D ultrasound","feed_subtitle":"Best model: 86.62% single-view accuracy, 92.58% with two views, and 83.0% on synthetic ultrasound, above human experts","key_machinery":"The load-bearing object is the variational autoencoder (VAE) latent space built from spleen segmentations. A residual-block convolutional encoder maps a single coronal slice, or channel-wise concatenated coronal and transverse slices, to a 128-dimensional Gaussian latent code parameterised by mean and standard deviation; the decoder reconstructs the segmentation. Three volume estimators were tested on this code: nearest-neighbour search, linear regression on the latent mean, and an end-to-end regression VAE (RVAE) whose fully connected layers map the latent code to volume under a combined reconstruction/KL/regression loss. The RVAE is the variant that carries the reported accuracy, and its reparameterised sampling also generates confidence intervals. A second mechanism, the ultrasound semantic diffusion model (USDM), synthesises realistic ultrasound images from layouts made by rotating CT spleen segmentations and overlaying them on real ultrasound cones, providing the volume-labelled images needed to evaluate the full pipeline.","core_discovery":"The paper's central claim is that spleen volume is recoverable from the shape of the spleen as seen in standard 2D ultrasound views, without reconstructing a 3D volume. The authors train a residual-block convolutional variational autoencoder on 2D slices taken from 149 CT spleen segmentations, using the coronal slice with the largest cross-sectional area for the single-view setting and adding the corresponding transverse slice for the dual-view setting. Volume is read out from the 128-dimensional latent code; the best-performing variant, RVAE, adds a fully connected regression head to the latent space and trains it end-to-end with reconstruction, KL, and volume regression losses. The same framework can sample the latent distribution 100 times to produce 95% confidence intervals. To test the whole pipeline from ultrasound rather than from clean segmentations, the authors generate synthetic ultrasound images from CT-derived segmentations with a semantic diffusion model conditioned on US layouts; a U-Net segments these images and the RVAE estimates volume, yielding an overall MRVA of 83.0%, above the human expert baseline of 68.54%.","pith_inferences":["If paired real ultrasound and CT/MRI volumes become available, the same architecture would provide a direct test of whether the 83.0% synthetic-pipeline accuracy transfers to patients; the paper does not claim this transfer has been shown.","Because the VAE regresses volume from shape rather than pixel texture, the method may extend to other organs whose volume matters clinically but where 3D imaging is impractical, such as the liver or kidneys, although each organ would need its own latent space and validation.","The authors' adult-CT versus paediatric-US mismatch suggests age-stratified calibration may be needed before clinical deployment, since spleen size and growth norms differ across childhood."],"forward_implications":["Spleen volume could be measured in routine 2D ultrasound examinations, removing the need for CT or MRI when assessing splenomegaly in settings where 3D imaging is unavailable.","A single coronal view suffices for useful volume estimates (86.62% MRVA on CT-derived segmentations), and adding a transverse view improves accuracy to 92.58%.","Confidence intervals and latent-space visualisation give clinicians a way to judge and trust individual volume estimates rather than receiving a bare scalar.","The diffusion-generated synthetic ultrasound dataset with ground-truth volumes is released publicly, giving other groups paired image-volume data that did not exist before.","Because the pipeline is built from standard clinical views, it could be integrated into existing ultrasound workflows for spleen assessment without changing acquisition practice."],"supporting_citations":[{"why":"Supplies 60 of the 149 CT spleen segmentations with manual labels used to train and test the volume estimation models.","marker":"Antonelli et al., 2022"},{"why":"Supplies 89 CT spleen segmentations with manual labels, the other main source of ground-truth volumes for training and evaluation.","marker":"Gibson et al., 2018"},{"why":"Provides the U-Net spleen segmentation model and its hyperparameter settings that the DeepSPV pipeline reuses for the segmentation stage.","marker":"Yuan et al., 2022"},{"why":"Provides the semantic diffusion backbone that the ultrasound semantic diffusion model adapts for synthetic ultrasound image generation.","marker":"Wang et al., 2022"},{"why":"Defines the Pix2Vox++ 2D-to-3D reconstruction baseline that DeepSPV is compared against for volume estimation.","marker":"Xie et al., 2020"},{"why":"Provides the single-measurement spleen length regression formula used as the human expert baseline and as part of the synthetic US evaluation comparison.","marker":"Bezerra et al., 2005"},{"why":"Provides the three-measurement linear regression formula for spleen volume used as the dual-view clinical baseline.","marker":"Prassopoulos et al., 1997"},{"why":"Supplies the interpretability approach of visualising a VAE latent space with PCA and decoding samples, which DeepSPV adapts to spleen segmentations.","marker":"Puyol-Antón et al., 2020a"}],"fun_headline_variants":["Deep learning estimates spleen volume from 2D ultrasound","Spleen volume from 2D ultrasound: deep learning beats human experts","First deep learning pipeline for 3D spleen volume from 2D ultrasound","Deep learning predicts spleen volume from 2D ultrasound alone","2D ultrasound to 3D spleen volume via deep learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The full-pipeline accuracy of 83.0% assumes that synthetic ultrasound images made by rotating and manually placing CT spleen segmentations into real ultrasound cones are representative enough of real clinical 2D ultrasound for the measured volume accuracy to carry over to patients, and that the coronal CT slice with the largest cross-sectional area matches the clinically recorded coronal view; no paired real ultrasound with volumetric ground truth is used anywhere in the evaluation.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning estimates spleen volume from 2D ultrasound","Spleen volume from 2D ultrasound: deep learning beats human experts","First deep learning pipeline for 3D spleen volume from 2D ultrasound","Deep learning predicts spleen volume from 2D ultrasound alone","2D ultrasound to 3D spleen volume via deep learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001226,"raw_usage":{"total_tokens":5112,"prompt_tokens":1090,"completion_tokens":4022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":706,"completion_tokens_details":{"reasoning_tokens":3933}},"tokens_in":706,"tokens_out":4022,"duration_ms":24559,"temperature":1.0,"reasoning_tokens":3933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:49:13.817894+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect paired 2D ultrasound images and CT or MRI spleen volumes from the same patients, run DeepSPV on the real ultrasound images, and compare mean relative volume accuracy with the CT volumes; if the accuracy falls substantially below the reported 83% or below the human-expert baseline, the synthetic-to-real transfer assumption fails.","supporting_citations":[{"cited_title":", author Giganti, F","cited_arxiv_id":null,"evidence_quote":"Supplies 89 CT spleen segmentations with manual labels, the other main source of ground-truth volumes for training and evaluation."},{"cited_title":", author Puyol-Ant \\'o n, E","cited_arxiv_id":null,"evidence_quote":"Provides the U-Net spleen segmentation model and its hyperparameter settings that the DeepSPV pipeline reuses for the segmentation stage."},{"cited_title":", author Yao, H","cited_arxiv_id":null,"evidence_quote":"Defines the Pix2Vox++ 2D-to-3D reconstruction baseline that DeepSPV is compared against for volume estimation."},{"cited_title":", author Daskalogiannaki, M","cited_arxiv_id":null,"evidence_quote":"Provides the three-measurement linear regression formula for spleen volume used as the dual-view clinical baseline."}],"review_version":1}