{"id":"b40e8e42-3d44-4dbb-817a-658889208aa0","arxiv_id":"2411.15322","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A deep learning pipeline can automatically outline the liver dome in kV-triggered images with mean errors of 6.4 to 7.7 mm and detection rates of 92% and 76% in two patient folds.","lead":"This paper trains a U-Net to automatically outline the liver dome in X-ray images taken during breath-hold liver radiation therapy. The system runs in under a second per image, with average positioning errors of 6 to 8 millimeters, and is a step toward automatic breath-hold verification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's headline accuracy numbers are uninterpretable without a precise RMSE definition: the Evaluation section never states whether undetected domes are excluded from RMSE or how contour correspondence is established. The 6.4–7.7 mm figures therefore cannot be independently verified.","rationale":"The reader's conditional verdict identifies the manual-contour ground-truth assumption as the weakest point, and that is indeed a legitimate concern. However, the more load-bearing issue is that the primary quantitative outcome—the RMSE between predicted and manual dome contours—is not defined well enough to be checked at all. Without knowing whether failed detections are excluded, how point correspondence is established, or how pixel distances are converted to millimeters, the headline numbers 6.4 mm and 7.7 mm cannot be interpreted or reproduced. This is not just a missing detail; it directly undermines the strongest claim that the pipeline is 'reasonably accurate.' The same criticism applies to the unstated morphological kernels, which further blocks independent reproducibility. These are addressable clarifications rather than evidence of a wrong result, so the appropriate verdict remains conditional: the proof-of-principle could be valid, but the manuscript as written does not permit verification of its central metric. Therefore the reader's CONDITIONAL verdict is retained, with the rationale shifted from ground-truth uncertainty to metric under-specification.","tokens_in":10129,"tokens_out":4058,"duration_ms":39178,"concrete_test":"Request the exact RMSE computation procedure from the authors: (1) specify whether undetected-dome images are excluded; (2) provide the contour-correspondence/resampling algorithm; (3) provide the pixel-to-mm calibration used for each image or a global conversion. Then recompute the Fold 2 mean RMSE under the stated rule, including all 345 images (assigning a large penalty or marking undetected cases as missing). If the recalculated all-image RMSE differs substantially from 7.7 mm, or if the procedure cannot be stated precisely, the abstract's accuracy figures should be revised or qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a fast, reasonably accurate automatic liver dome delineation is achievable—rests on the reported RMSE values (6.4±1.6 mm in Fold 1, 7.7±2.3 mm in Fold 2) and detection rates (91.7% and 76.3%). The Evaluation section defines the second metric only as 'divergence ... quantified using the Root Mean Square Error (RMSE) values between the predicted and ground truth contours,' but never specifies: (i) whether images in which postprocessing produces a non-existent or too-sparse dome are included in the RMSE average, and if so, what value is assigned; (ii) how the predicted contour and manual contour are matched or resampled for the pointwise RMSE computation; or (iii) what physical pixel spacing or magnification factor converts image pixels to millimeters. If undetected images are excluded, then in Fold 2 the RMSE is computed on only 76.3% of the images, and the true accuracy over all triggered images is unknown and likely worse; in Fold 1 the exclusion affects 8.3% of images. The proof-of-principle accuracy claim therefore depends on an undefined evaluation protocol. In addition, the postprocessing step uses 'custom kernels' that are not described, so even with a defined RMSE the numbers cannot be reproduced independently. This concern is more direct than ground-truth bias: even if the manual contours and 20-pixel expansion are perfect, the reported RMSE is not verifiable without the exact evaluation procedure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This proof-of-principle study proposes an automatic pipeline for delineating the liver dome in kV-triggered planar images acquired during breath-hold verification for liver SBRT. The pipeline consists of a U-Net with a VGG16 encoder pretrained on ImageNet to segment a liver region, followed by thresholding, largest-connected-component extraction, HED edge detection, Canny edge detection, and morphological operations with custom kernels to extract the liver dome contour. The dataset comprises 711 kV-triggered images from 24 patients, split into two patient-level folds for two-fold cross-validation. The U-Net is trained on augmented data (500 images per fold) with early stopping. The authors report IoU > 0.9 for liver region segmentation for all patients, and for the final pipeline, an RMSE of 6.4±1.6 mm (Fold 1) and 7.7±2.3 mm (Fold 2), with detection rates of 91.7% and 76.3%, respectively. Training takes about 30 minutes and inference under one second. The conclusion is that the pipeline is a clinically feasible proof-of-principle for automatic liver dome delineation.","tokens_in":10438,"tokens_out":6063,"duration_ms":47384,"significance":"If the reported accuracy is reproducible, the study makes a useful contribution: it is, to the authors' knowledge, the first DL-based liver dome delineation from kV planar images, and it demonstrates a novel use of HED edge detection in postprocessing. The patient-level two-fold cross-validation is an appropriate design for generalization assessment, and the reported IoU values are consistently high. The fast inference time and modest training resource requirement are strengths for potential online clinical use. However, the paper's headline accuracy figures (RMSE and detection rate) rest on an incompletely specified evaluation protocol and an undescribed postprocessing step, so the central quantitative claims cannot currently be independently verified. The ground-truth mask construction also relies on an unvalidated 20-pixel inferior expansion. These issues are fixable and do not invalidate the feasibility concept, but they do prevent acceptance in the present form.","major_comments":[{"comment":"The definition of the RMSE metric is incomplete. The text (Evaluation section, p. 14) states only that RMSE is computed 'between the predicted and ground truth contours,' without specifying (i) whether images in which the postprocessing yields no dome or a too-sparse dome are excluded from the RMSE average, and what value is assigned if they are included; (ii) how the predicted contour is matched to the manual contour or how points are resampled for the pointwise computation; and (iii) the physical pixel spacing or magnification factor used to convert pixel distances to millimeters. Because detection rates are 91.7% and 76.3%, the RMSE may be computed on a favorable subset of images, which would not represent the accuracy on all triggered images. These details must be provided before the reported 6.4±1.6 mm and 7.7±2.3 mm values can be interpreted.","section":"Evaluation"},{"comment":"The postprocessing step is not reproducible as described. The paper (Postprocessing, p. 14) refers to 'morphological postprocessing with custom kernels' and 'erosion operations with custom kernels' without specifying the kernel sizes, shapes, number of iterations, or the interpolation method used to connect contour discontinuities. Without this information, the final dome extraction cannot be replicated by another group, and the reported RMSE and detection rates are not independently verifiable. Please provide the exact morphological recipe and algorithm.","section":"Postprocessing"},{"comment":"The training masks for the U-Net are created by expanding the manual liver dome contours inferiorly by 20 pixels (Image preprocessing and ground truth labeling, p. 10). This expansion is asserted to form a 'liver region' mask, but no justification or sensitivity analysis is provided. If the 20-pixel expansion does not match the true liver region (for example, because it encroaches on other anatomy or is too narrow), both the IoU scores and the contours extracted in postprocessing—and hence the final RMSE—are affected. Please validate this assumption, e.g., by comparing the expanded masks with an independent liver region reference or by varying the expansion width and reporting the impact on final metrics.","section":"Image preprocessing and ground truth labeling"},{"comment":"The detection-rate metric is not fully specified. The Evaluation section (p. 14) defines a dome as undetectable when the 'resulting dome contour ... is non-existent or too sparse,' but 'too sparse' is never quantified. This subjective threshold can materially change the reported 91.7% and 76.3% detection rates. Please define an operational criterion (e.g., minimum number of contour points, minimum contour length, or minimum fraction of the ground-truth contour length).","section":"Evaluation"}],"minor_comments":[{"comment":"The figure caption reads 'Postprocessing seps' but should read 'Postprocessing steps'.","section":"Figure 4"},{"comment":"In the Discussion (p. 19), the sentence 'In both cases, good learning and generalizing of the models can be observed because the validation learning curves remain higher than the training curves' is unclear: a validation loss that remains above the training loss is typically a sign of some overfitting, not necessarily 'good' generalization. Please clarify the intended interpretation.","section":"Discussion"},{"comment":"In the Results (p. 17), the statement 'For patients 1 to 12, comprising Fold1, the mean divergence ranged from 4.1 mm to 9.6 mm' does not specify which model (trained on Fold2 or Fold1) produced these values; clarify the fold-to-model mapping.","section":"Results"},{"comment":"Figure 7(a) marks outliers with circles, but no formal outlier definition is given in the text.","section":"Figure 7(a)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and the proof-of-principle idea is timely. The main concern is the incomplete evaluation protocol, which is fixable in revision. No concerns about data provenance or ethics. The authors should be asked to provide the missing definitional and algorithmic details before the paper is considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious look: it is, to my knowledge, the first DL pipeline for liver dome delineation on kV planar images for breath-hold verification, and the authors are appropriately modest about its clinical readiness. The design is reasonable — a U-Net with a frozen VGG16 encoder, HED edge maps as an extra input channel and again in postprocessing, followed by thresholding and morphology. The evaluation is patient-level two-fold cross-validation, the IoU values for the liver-region segmentation are consistently above 0.9, and the pipeline is fast enough for online use. The paper also honestly flags small dataset size, the lack of a hold-out validation set, and the augmentation limits. These are real strengths.\n\nThe soft spots are mostly about reproducibility of the reported numbers. The RMSE metric is never actually defined: the paper says divergence is quantified as RMSE between predicted and ground truth contours, but it does not say whether undetected domes are excluded, how the contours are matched or resampled, or what pixel spacing is used. If undetected images are dropped, the Fold 2 RMSE is computed on only 76% of the images, and the true accuracy over all triggered images is unknown. That is a load-bearing issue for the abstract's central claim. The custom morphological kernels are also not described, so even with a defined metric the numbers cannot be reproduced from the text. There is no baseline comparison, and no code or data. These are all addressable in revision.\n\nI do not think the ground-truth concern is the main problem. The manual contours from the prior study are a reasonable reference, and the 20-pixel inferior expansion is an explicit choice, even if somewhat ad hoc. The undefined RMSE is more direct.\n\nWho is this for? Medical physicists working on image-guided liver SBRT or on automated verification of breath-hold reproducibility. It deserves a proper peer review — the clinical motivation is clear, the approach is novel for this modality, and the flaws are fixable with more precise methodology reporting. I would not desk-reject it, but I would send it back with the expectation that the authors define the evaluation metrics precisely and ideally release enough detail to reproduce the pipeline. My own verdict is skeptical on the exact numbers, not on the underlying feasibility.","headline":"A sensible proof-of-principle for an underexplored clinical task, but the headline RMSE numbers are not defined tightly enough to verify without clarification.","tokens_in":10988,"tokens_out":1939,"would_cite":false,"duration_ms":19498,"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 deep-learning pipeline can delineate the liver dome on kV triggered images in under one second, with a reported error of 6.4–7.7 mm RMSE against manual contours and detection rates of 92% and 76% in two patient folds.","keywords":["liver SBRT","breath-hold reproducibility","kV triggered imaging","liver dome delineation","deep learning segmentation","U-Net","HED edge detection","online image guidance"],"falsifier":"Have a second radiation oncologist independently re-contour the liver domes on the same 711 images. If the inter-observer RMSE between the two manual contour sets is comparable to or larger than the 6.4–7.7 mm RMSE reported here, then the pipeline's agreement with ground truth is within the noise of manual delineation rather than an independent measure of anatomical accuracy.","tokens_in":9902,"feed_emoji":"🩻","tokens_out":8705,"duration_ms":67739,"temperature":0.7,"pith_summary":"The paper aims to show that a deep-learning pipeline can replace manual visual checks of liver dome position during breath-hold liver SBRT. It assembles a U-Net with transfer learning and a pretrained edge-detection model to segment and then trace the liver dome on 711 kV-triggered images from 24 patients. Reported results are fast (under one second per image) and moderately accurate (6.4–7.7 mm RMSE versus manual contours, with detection rates of 92% and 76% in the two cross-validation folds). The authors position the work as a proof of principle, not yet a clinically validated tool, and say further improvement in accuracy and detection rate is needed before clinical use. If the approach holds up, it would enable automatic, objective breath-hold verification and eventually automatic beam gating in liver SBRT.","feed_headline":"AI pipeline delineates liver dome in under one second per image","feed_subtitle":"Breath-hold checks for liver SBRT could become automatic: reported errors are 6.4-7.7 mm versus manual contours.","key_machinery":"The load-bearing mechanism is the combination of a U-Net (an encoder-decoder convolutional network with skip connections, here with a frozen pretrained encoder) that predicts a liver-dome-region mask, and a hand-designed postprocessing chain that converts the mask into a dome contour: binary thresholding at mean intensity, largest connected component, HED edge detection, edge thinning, and morphological erosion and dilation with custom kernels to isolate the upper boundary. A three-channel input (original image, contrast-enhanced image, and HED edge map) supplies the network with the soft-tissue contrast and edge information that planar kV images lack.","core_discovery":"The paper claims that a pipeline consisting of a U-Net trained to segment the liver dome region, followed by thresholding, extraction of the largest connected component, HED edge detection, edge thinning, and morphological filtering, can automatically delineate the liver dome on planar kV-triggered images. In two-fold cross-validation on 711 images from 24 patients, the model's liver-region predictions reached IoU above 0.9 for every patient, the end-to-end delineation ran in under one second per image, and the extracted dome contours differed from manual contours by 6.4±1.6 mm (Fold 1, detection rate 91.7%) and 7.7±2.3 mm (Fold 2, detection rate 76.3%).","pith_inferences":["If the same pipeline were trained on diaphragm or other organ-at-risk masks, it could extend to breath-hold verification for lung or abdominal SBRT, since the method itself is not liver-specific.","The paper's observation that augmentation beyond 500 images degraded performance suggests the synthetic transformations did not match real image variation; richer augmentation or multi-institutional data might improve the lower detection rate in Fold 2.","Combining the predicted dome contour with the planning-CT dome position would enable automatic beam gating, which the paper names as the eventual clinical goal.","A failure analysis separating segmentation errors from postprocessing errors, as hinted at for patients 03, 14, and 19, could pinpoint whether the 76% detection rate in Fold 2 is due to anatomy, image quality, or the morphological step."],"forward_implications":["Automatic liver dome delineation in under one second per image makes real-time online breath-hold verification feasible during treatment.","Replacing manual visual comparison with an automatic contour could reduce subjectivity and human intervention in breath-hold checks.","The reported 6.4–7.7 mm RMSE is within the range needed to catch breath-holds with more than 10 mm deviation, but the authors state that the accuracy and detection rate must improve before the technique is clinically relevant.","Two-fold cross-validation with patient-level splits suggests the pipeline generalizes across patients at the same institution, while multi-institutional testing remains an open step.","The pipeline could serve as the baseline for the first deep-learning-based liver dome delineation on kV planar images."],"supporting_citations":[{"why":"Supplies the kV-triggered image dataset and the manual liver dome contours used as ground truth, and defines the breath-hold verification procedure this pipeline automates.","marker":"[1]"},{"why":"Provides pretrained natural-image weights for transfer learning of the U-Net encoder.","marker":"[20]"},{"why":"Provides the HED edge-detection model used both to build the network's edge-map input channel and to extract edges from predicted liver regions in postprocessing.","marker":"[22]"},{"why":"Defines CLAHE, the contrast-enhancement step used to generate the first input channel for low-contrast kV images.","marker":"[24]"},{"why":"Defines the U-Net encoder-decoder architecture with skip connections that the segmentation stage is built on.","marker":"[34]"}],"fun_headline_variants":["AI delineates liver dome in under a second for SBRT","Automatic liver dome contouring: 1 second, ~7 mm accuracy","Deep learning speeds liver dome checks for SBRT breath-hold","U-Net pipeline auto-delineates liver dome in <1 sec per image","Fast AI liver dome detection for breath-hold verification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The manual liver dome contours from the earlier study are accurate enough to serve as ground truth, and expanding each one inferiorly by 20 pixels yields a valid liver-region training mask; if either fails, the IoU scores and RMSE values lose their meaning.","fun_headline_variants_meta":{"raw":{"variants":["AI delineates liver dome in under a second for SBRT","Automatic liver dome contouring: 1 second, ~7 mm accuracy","Deep learning speeds liver dome checks for SBRT breath-hold","U-Net pipeline auto-delineates liver dome in <1 sec per image","Fast AI liver dome detection for breath-hold verification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1392,"prompt_tokens":1012,"completion_tokens":380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":628,"tokens_out":380,"duration_ms":4109,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:25:06.304249+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have a second radiation oncologist independently re-contour the liver domes on the same 711 images. If the inter-observer RMSE between the two manual contour sets is comparable to or larger than the 6.4–7.7 mm RMSE reported here, then the pipeline's agreement with ground truth is within the noise of manual delineation rather than an independent measure of anatomical accuracy.","supporting_citations":[{"cited_title":"Online verification of breath- hold reproducibility using kV-triggered imaging for liver stereotactic body radiation therapy","cited_arxiv_id":null,"evidence_quote":"Supplies the kV-triggered image dataset and the manual liver dome contours used as ground truth, and defines the breath-hold verification procedure this pipeline automates."},{"cited_title":"Accessed November 10, 2023","cited_arxiv_id":null,"evidence_quote":"Provides pretrained natural-image weights for transfer learning of the U-Net encoder."},{"cited_title":"Accessed November 10, 2023","cited_arxiv_id":null,"evidence_quote":"Defines CLAHE, the contrast-enhancement step used to generate the first input channel for low-contrast kV images."},{"cited_title":"U-Net: Convolutional Networks for Biomedical Image Segmentation","cited_arxiv_id":null,"evidence_quote":"Defines the U-Net encoder-decoder architecture with skip connections that the segmentation stage is built on."}],"review_version":1}