{"id":"9abae6db-4e9a-4b05-8d79-d28963ba2d4e","arxiv_id":"2504.21231","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"T2ID-CAS adds SDXL-generated synthetic tracheal ring and vocal fold images plus class-aware sampling to raise neck ultrasound landmark detection mAP50-95 from 66 to 88.2.","lead":"This paper combines a text-to-image diffusion model with class-aware sampling to create synthetic ultrasound images that balance underrepresented classes, improving YOLOv9 landmark detection from 66 to 88.2 mean average precision. A generalist reader might care because it shows a practical route to fix class imbalance in medical image datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic-image annotation is unspecified: the entire T2ID-CAS mAP gain hinges on how 600 diffusion outputs became YOLO bounding-box labels, which the paper never states.","rationale":"Good-faith reading: the paper is internally coherent, and the arithmetic in Table II is consistent: per-class numbers average to the reported all-class mAP. The standalone CAS (84.3) already accounts for much of the gain over baseline (66); the additional claim is that SDXL synthetic data adds further gain (84.3 to 88.2) and lifts tracheal ring from 63.4 (CAS) to 90.5. That additional claim is exactly the point at which the experimental protocol is underspecified. Detection training cannot consume raw generated images without labels, yet the paper only describes annotation for the real scans. This is not an out-of-consensus disagreement; it is an internal completeness gap. The same gap applies to whether synthetic samples were derived from images that also served as validation data. Both conditions are necessary for the experiment to have the meaning claimed. They are also easily checked, so the reader's CONDITIONAL verdict is appropriate; if the authors cannot supply these details or the re-check fails, the mAP claim would need to be downgraded.","tokens_in":7805,"tokens_out":5015,"duration_ms":52147,"concrete_test":"Ask the authors to release one complete training sample for the synthetic branch: one generated synthetic image and its corresponding YOLO label file (txt). Then run a controlled re-training of YOLOv9s under the T2ID-CAS protocol with two modifications: (a) use only per-instance boxes that an airway specialist drew on the synthetic images, not image-level or pseudo-labels; and (b) exclude from SDXL fine-tuning any images whose subjects appear in the validation fold. Compare per-class mAP50-95 to Table II. If tracheal-ring mAP drops by more than 5 points under either modification, the reported 88.2 is not supported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reported 88.2 mAP (Table II) is driven by synthetic tracheal-ring and vocal-fold images, but the paper never describes how these 600 generated images were labeled. Section III.C says only that 600 images of size 512x512 were generated, 300 per class; the sole annotation description in Section III.A concerns real US scans (\"annotated by airway specialists using YoloMark\"). For YOLOv9s training, every image requires a per-instance bounding-box label file. If synthetic images are labeled with whole-image boxes, image-level labels, or pseudo-labels from an unstated detector, then the large tracheal-ring improvement (38.5 to 90.5) may reflect texture/class cues rather than genuine anatomical localization. A second unstated premise is data hygiene: 840 real images are used to fine-tune SDXL-LoRA, yet the paper does not say whether these overlap the 4-fold cross-validation folds. If the same subjects/images that appear in the validation fold were used in diffusion fine-tuning, the generated synthetic samples can memorize validation anatomy and inflate the comparison. Without either detail, Table II and the central claim are not reproducible. (As a separate reporting error, Table I lists SD v1-4 tracheal-ring IS as 11.045 while the text says 17.045; this does not change the main concern but reinforces the need for a code/data release.)","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes T2ID-CAS, a hybrid approach for mitigating class imbalance in neck ultrasound anatomical landmark detection. The method combines class-aware sampling (CAS) with synthetic image generation from a fine-tuned Stable Diffusion XL (SDXL) model using Low-Rank Adaptation (LoRA). The authors generate 600 synthetic images for the two most underrepresented classes (tracheal rings and vocal folds), augment the real training data with them, and train a YOLOv9s detector. They report a mean Average Precision (mAP50-95) of 88.2 with the proposed approach, compared with a baseline of 66, and substantial per-class gains for the minority classes. The paper also evaluates alternative strategies (mosaic/mixup, repeat factor sampling, and their combinations) and reports generative image quality metrics (FID, IS, CLIP score) for the synthetic images.","tokens_in":8096,"tokens_out":4748,"duration_ms":40615,"significance":"If the reported results are reproducible and the methodology is sound, the paper offers a practical and computationally efficient strategy for handling class imbalance in small medical ultrasound datasets, a setting where standard oversampling and GAN-based approaches often struggle. The integration of parameter-efficient diffusion fine-tuning (LoRA) with class-aware sampling is a reasonable and timely contribution. The paper also provides quantitative evidence that synthetic data can substantially improve minority-class detection, which is clinically relevant for airway management. However, the significance is currently limited by missing methodological details that are essential for reproducing and interpreting the central claim.","major_comments":[{"comment":"The paper never describes how the 600 synthetic images are annotated with bounding boxes for YOLOv9 training. Section III.A specifies annotation only for real US scans ('annotated by airway specialists using YoloMark'), while Section III.C only states that 600 synthetic images of size 512x512 were generated, 300 per class. Since YOLOv9 training requires per-image bounding-box label files, the reader cannot know whether the synthetic boxes are whole-image boxes, image-level labels, or pseudo-labels from an unstated detector. This detail is load-bearing because the reported mAP improvement for tracheal ring (38.5 to 90.5) and vocal fold (75.6 to 98.2) could be inflated if the synthetic labels do not represent genuine anatomical localization. Please specify the annotation protocol for synthetic images and, if applicable, how box coordinates were assigned.","section":"III.C"},{"comment":"The paper does not state whether the 840 real images used to fine-tune SDXL-LoRA are disjoint from the validation folds in the 4-fold cross-validation. If the same subjects or images appear both in the diffusion fine-tuning set and in a validation fold, the synthetic images can memorize validation-level anatomy and inflate the reported mAP. The paper only says 'The dataset was split into training and validation sets, with 4-fold cross-validation' but does not describe the split granularity (subject-level versus image-level) or confirm that the fine-tuning set was excluded from validation. Please clarify the exact split and provide a leakage check, such as ensuring no subject overlap between the SDXL fine-tuning set and any validation fold.","section":"III.B"},{"comment":"The cross-validation results in Table II are reported only as aggregate mAP50-95, with no per-fold results, standard deviations, or statistical significance tests. Given the small number of subjects (10) and the long-tailed class distribution, the observed differences (e.g., 88.2 for T2ID-CAS versus 84.3 for CAS alone, or 66 for baseline) may not be robust across folds. Without error bars or fold-wise numbers, the reader cannot assess the reliability of the claimed improvement. Please report fold-level mAP for all configurations, or provide confidence intervals and, if feasible, a paired significance test.","section":"IV.B"},{"comment":"There is an inconsistency in the reported Inception Score for the SD v1-4 tracheal ring condition: the text in Section IV.A states 'notable improvements in IS (18.184 vs. 17.045)' while Table I lists the SD v1-4 tracheal ring IS as 11.045. This discrepancy affects the interpretation of the generative quality comparison and must be corrected. Please verify all numbers in Table I against the text and any supplementary material.","section":"IV.A / Table I"}],"minor_comments":[{"comment":"The description of the 4-fold cross-validation split is ambiguous: it should state explicitly whether the split is at the subject level or the image level, since cineloop frames from the same subject are highly correlated and image-level splits can overestimate generalization.","section":"III.B"},{"comment":"The prompts shown in Fig. 3 ('Ultrasound image of human tracheal ring') differ from those reported in Section III.B ('High-resolution ultrasound scan of the human tracheal ring'). Please unify the prompt text across the paper.","section":"Fig. 3 / III.B"},{"comment":"The comparison between SDXL-LoRA and SD v1-4 is not controlled, since SD v1-4 is not fine-tuned on the ultrasound data; please state this explicitly and, if possible, include a fine-tuned SD v1-4 baseline for a fairer comparison of model architectures.","section":"IV.A"},{"comment":"It would be helpful to report the number of synthetic images used per class in the final training set and the resulting class distribution, so that the effect of the augmentation is transparent.","section":"Table II"},{"comment":"The paper does not mention code or data availability; providing at least the code for the T2ID-CAS pipeline and the synthetic-data generation process would substantially improve reproducibility.","section":"General"},{"comment":"There are minor typographical errors, e.g., 'SDXL synthetic image images' in Section IV.B and the rendering of 'Frécet' as 'Fr ´echet'; these should be corrected during revision.","section":"IV.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is concise and the core idea is relevant, but the missing details on synthetic-image annotation and on leakage control between the SDXL fine-tuning set and the cross-validation folds are essential for the central claim. These issues are addressable by the authors, so I recommend major revision rather than rejection. The IS inconsistency and the absence of error bars also need attention before the paper can be considered for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. T2ID-CAS is a plausible, incremental engineering contribution: fine-tune SDXL-LoRA on neck-ultrasound images of two minority classes, generate 600 synthetic images, combine with class-aware sampling, and retrain YOLOv9s. The reported jump from 66 to 88.2 mAP50-95 is internally consistent with the per-class table. The paper does some things well: it compares several imbalance strategies (mosaic/mixup, RFS, CAS, with and without synthetic data), reports FID/IS/CLIP for the generator, uses 4-fold cross-validation, and includes an honest limitations paragraph. The per-class gains for tracheal ring (38.5 to 90.5) and vocal folds (75.6 to 98.2) are large, and if they are real, the method is practically useful.\n\nNow the soft spots, in rough order of importance. First, the paper never explains how the 600 synthetic images were labeled with bounding boxes. Section III.A describes annotation only for real scans; Section III.C says only that images were generated and resized. YOLOv9 training needs per-instance box files. If the synthetic images are labeled with whole-image boxes or pseudo-labels from some unstated detector, the tracheal-ring and vocal-fold improvements could reflect texture/class cues rather than anatomical localization. Second, the SDXL-LoRA fine-tuning used 840 real images, but the paper does not say whether those overlap the 4-fold cross-validation folds. If validation subjects were included in diffusion fine-tuning, synthetic images could memorize validation anatomy and leak. I want to be clear: there is no evidence of leakage, and the burden is low, but the omission blocks reproducibility. Third, there is a numerical inconsistency in Table I: the text reports 17.045 for SD v1-4 tracheal-ring IS, while the table lists 11.045. Fourth, there are no error bars, and the fold split strategy is unspecified. These are all fixable, and I would not call the central claim circular: the mAP gain is a train-test measurement, not the output of a fitted equation.\n\nWho is this for? People working on class imbalance in medical ultrasound object detection, and anyone who wants a cautionary case study on synthetic-data reporting. It is an incremental combination of published components, not a new mechanism. It deserves a serious referee because the reported effect size is large and the gaps are addressable. I would send it out, but I would ask for the annotation details, data hygiene statement, error bars, and a corrected table before accepting.","headline":"A plausible but incompletely reported diffusion-plus-class-aware-sampling pipeline for neck-ultrasound landmark detection; the large mAP gain rests on unstated synthetic-image labeling and possible data overlap.","tokens_in":8587,"tokens_out":2067,"would_cite":false,"duration_ms":20518,"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":"Fine-tuning SDXL on minority-class ultrasound frames and balancing batches with class-aware sampling lifts YOLOv9s neck-landmark mAP50-95 from 66 to 88.2.","keywords":["neck ultrasound","anatomical landmark detection","class imbalance","text-to-image diffusion","SDXL","class-aware sampling","YOLOv9","synthetic data augmentation"],"falsifier":"Re-run the experiment with specialist-verified bounding boxes on all 600 synthetic images and with an audit confirming that none of the 840 fine-tuning images overlap the validation folds; if the tracheal-ring mAP gain over CAS alone then collapses below the baseline difference, the central claim is unsupported.","tokens_in":7649,"feed_emoji":"🩺","tokens_out":4608,"duration_ms":46005,"temperature":0.7,"pith_summary":"The paper claims that class imbalance in neck-ultrasound anatomical landmark detection can be substantially mitigated by combining two ingredients: synthetic images of underrepresented structures produced by a fine-tuned text-to-image diffusion model, and class-aware sampling that builds balanced training batches. On a six-class dataset from ten subjects, the hybrid T2ID-CAS pipeline raises YOLOv9s mean average precision (mAP50-95) from 66 to 88.2, with tracheal-ring detection climbing from 38.5 to 90.5 and vocal folds reaching 98.2. The authors argue this shows diffusion-based augmentation, rarely applied to ultrasound, is a computationally efficient and scalable way to make detection models usable for underrepresented but clinically important airway structures.","feed_headline":"Synthetic ultrasound images raise airway-landmark detection from 66 to 88.2 mAP","feed_subtitle":"Class-aware sampling plus diffusion data pushes tracheal-ring detection from 38.5 to 90.5 in neck ultrasound.","key_machinery":"The load-bearing mechanism is the combination of SDXL fine-tuned with LoRA on class-specific ultrasound images to generate high-fidelity synthetic minority-class images, and class-aware sampling, which maintains a list of classes and randomly picks an image containing a selected class per minibatch to guarantee uniform representation of every anatomical class. A YOLOv9s detector, using Programmable Gradient Information and the GELAN backbone, is then trained on the hybrid real-plus-synthetic dataset. The synthetic images supply the underrepresented classes with extra annotated instances, while CAS ensures those instances are actually seen in each batch; together they move tracheal-ring mAP from roughly 38 to 90, whereas either mechanism alone gives far less for that class.","core_discovery":"T2ID-CAS fine-tunes SDXL, a latent text-to-image diffusion model, on 840 real ultrasound images of tracheal rings and vocal folds using Low-Rank Adaptation (LoRA), then generates 600 synthetic images (300 per class) conditioned on anatomical text prompts. These synthetic samples are combined with the real training set, and a class-aware sampler selects images during training so each anatomical class appears in minibatches at a uniform rate. On YOLOv9s with 4-fold cross-validation, the full pipeline achieves an mAP50-95 of 88.2 versus 66 for the baseline, while CAS alone reaches 84.3; per-class gains are largest for the minority classes, and the authors note some well-represented classes such as thyroid and cricoid cartilage show minor regression when synthetic data are added.","pith_inferences":["If the synthetic labels were derived by reusing real-image annotations rather than re-annotating each generated image, the generation step could be leaking instance-level information; this is testable by inspecting the labeling pipeline, which the paper does not describe.","The same pipeline could transfer to other ultrasound tasks, such as fetal or vascular landmark detection, but the prompt sensitivity the authors mention suggests gains may vary with how precisely the text prompt matches the target anatomy.","A stronger comparison would hold compute constant: the fine-tuning and generation cost of SDXL could instead be spent on more real data or extended training, a trade-off the paper does not quantify.","An independent replication with strict separation between generator-training images and validation folds would clarify whether the reported gain is due to class balance or to accidental leakage."],"forward_implications":["If the reported numbers hold, diffusion-based augmentation plus class-aware sampling is an effective remedy for long-tailed medical detection datasets where minority structures are clinically critical.","Tracheal-ring detection, the weakest class at baseline, gains the most, implying the hybrid approach preferentially benefits classes with the most severe underrepresentation.","Because CAS alone accounts for most of the overall gain (84.3 of the 88.2 mAP), synthetic data acts as a refinement rather than the sole driver, suggesting both mechanisms should be used together.","Slight regression on majority classes like thyroid and cricoid cartilage indicates the trade-off of adding synthetic data to already well-sampled classes.","Comparisons with Repeat Factor Sampling show that not all class-imbalance remedies work equally with synthetic augmentation, since RFS variants stayed below 75 mAP even when combined with SDXL data."],"supporting_citations":[{"why":"Supplies the latent diffusion architecture that Stable Diffusion operates in, the base generation mechanism being fine-tuned.","marker":"[18]"},{"why":"Defines the SDXL model whose higher-resolution synthesis the paper fine-tunes with LoRA.","marker":"[21]"},{"why":"Provides Low-Rank Adaptation, the parameter-efficient fine-tuning method used to adapt SDXL to ultrasound images.","marker":"[22]"},{"why":"Source of the class-aware sample weighting idea the paper adapts as CAS.","marker":"[27]"},{"why":"Introduces YOLOv9, the object detector whose Programmable Gradient Information and GELAN backbone are used for landmark detection.","marker":"[31]"},{"why":"Establishes denoising diffusion probabilistic models, the generative framework underlying the text-to-image synthesis.","marker":"[13]"}],"fun_headline_variants":["Diffusion + class-aware sampling lifts neck US landmark mAP to 88.2","Synthetic data boosts minority-class detection in neck ultrasound","T2ID-CAS: diffusion model fixes class imbalance in neck US detection","Class imbalance? Diffusion-generated images fix neck US detection","From 66 to 88.2 mAP: diffusion tackles rare anatomy in ultrasound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported mAP gain holds only if the synthetic images are labeled with the same ground-truth boxes as real images and if no image used to tune the diffusion generator appears in the validation folds.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion + class-aware sampling lifts neck US landmark mAP to 88.2","Synthetic data boosts minority-class detection in neck ultrasound","T2ID-CAS: diffusion model fixes class imbalance in neck US detection","Class imbalance? Diffusion-generated images fix neck US detection","From 66 to 88.2 mAP: diffusion tackles rare anatomy in ultrasound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001334,"raw_usage":{"total_tokens":5401,"prompt_tokens":894,"completion_tokens":4507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":4412}},"tokens_in":510,"tokens_out":4507,"duration_ms":33286,"temperature":1.0,"reasoning_tokens":4412,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:08:31.294753+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the experiment with specialist-verified bounding boxes on all 600 synthetic images and with an audit confirming that none of the 840 fine-tuning images overlap the validation folds; if the tracheal-ring mAP gain over CAS alone then collapses below the baseline difference, the central claim is unsupported.","supporting_citations":[{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Establishes denoising diffusion probabilistic models, the generative framework underlying the text-to-image synthesis."}],"review_version":1}