{"id":"f3991503-e8e0-41ec-9864-752a0e2f6708","arxiv_id":"2508.13077","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A TTE-trained diffusion model can be adapted to generate realistic TEE images with small LoRA adapters, and these synthetic images improve downstream cardiac segmentation.","lead":"This paper adapts a diffusion model trained on transthoracic echocardiography (TTE) to generate transesophageal echocardiography (TEE) images using lightweight LoRA adapters and a mask-remapping layer called MaskR2. The authors show that mixing the generated synthetic TEE images with real ones improves multiclass cardiac segmentation, especially for right-heart structures.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Train/test split may not be video-exclusive; same-video frames across splits could inflate reported segmentation gains from synthetic TEE augmentation.","rationale":"The reader's stated weakest assumption is that TTE-pretrained features transfer to TEE via LoRA. That is a plausible concern, but the most load-bearing vulnerability in the present manuscript is the evaluation protocol for the downstream segmentation claim. The paper reports no patient/video-level split, no confidence intervals, and no per-video grouping, yet the central practical contribution is a measured Dice improvement on a small, internally collected TEE test set. If the 288 frames from 71 videos are split at the frame level, the same-video leakage alone could explain a substantial part of the improvement, independent of whether TTE features transfer. This concern is concrete, checkable from information the authors should already possess, and it directly determines whether the headline claim—synthetic TEE augmentation improves segmentation—is valid. The reader did flag missing error bars as a secondary issue, but not the video-exclusive split. Since the verdict is already CONDITIONAL, the additional required evidence (video-level split verification or a video-exclusive rerun) does not change the overall verdict; it sharpens the conditions under which acceptance would be justified.","tokens_in":7555,"tokens_out":7504,"duration_ms":78794,"concrete_test":"Request the video/patient identifiers for each of the 196/40/52 TEE frames (or release the code with the split). If any video appears in more than one split, rerun the nnUNet augmentation experiment with a video-exclusive split, e.g., 5-fold cross-validation at the video level, and recompute per-class Dice for real-only versus augmented training. If the right-heart Dice gains vanish or become statistically insignificant, the central claim is not supported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central downstream claim is that mixing under 200 real TEE frames with synthetic frames improves multiclass segmentation Dice, especially for right-heart structures. This claim rests on a comparison between nnUNet trained on real TEE frames plus synthetic images and nnUNet trained on real frames alone, evaluated on a 52-image test set. The TEE collection is described as 288 image–mask pairs drawn from 71 videos (about 4 frames per video), but the data split is described only as allocating 196/40/52 images 'ensuring an even distribution across views and cardiac phases.' There is no statement that all frames from a given video or patient are confined to a single split, nor any report of unique videos per split. If frames from the same video appear in both the training set (used both for diffusion adaptation and for nnUNet augmentation) and the test set, the reported Dice gains could be inflated by near-duplicate appearance: the generative and segmentation models could memorize patient-specific texture, speckle, probe geometry, and anatomy rather than learning a general cross-modality mapping. The manuscript also provides no error bars or per-video grouped statistics, so the magnitude of the improvement cannot be separated from this potential leakage. The absence of a video-exclusive split is a missing support that is directly load-bearing for the paper's headline application claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline that freezes an EDM diffusion model pretrained on CAMUS TTE mask-conditioned images and adapts it to TEE using LoRA adapters trained on only 196 real TEE image–mask pairs. A new MaskR2 layer remaps the TEE label set (LA, LV, RA, RV) into the base model's original TTE label space, so the conditioning architecture can be reused without retraining. The authors compare different LoRA adapter configurations on FID, LPIPS, and SSIM, and then train nnUNet segmentation models on real TEE frames mixed with synthetic images generated from either real masks or SSM-derived masks, reporting improvements in multiclass Dice, especially for right-heart structures. The abstract further claims that adapting only MLP layers suffices, while the methods state that cross-attention layers are always trained, an inconsistency that the paper does not resolve.","tokens_in":7862,"tokens_out":8882,"duration_ms":92648,"significance":"If the quantitative claims hold, this is a useful contribution to data-scarce echocardiography: it demonstrates parameter-efficient cross-modality adaptation with fewer than 200 real TEE pairs, proposes a simple MaskR2 label-remapping mechanism, and validates synthetic augmentation on a real downstream segmentation task rather than only on image-quality metrics. The use of external benchmarks (FID, LPIPS, downstream segmentation on real TEE) is a strength and avoids circularity. However, the version I reviewed does not establish data independence between training and test videos, and the internal contradiction about which layers are trained undermines the targeted-adaptation claim; both issues need to be addressed before the results can be considered reliable.","major_comments":[{"comment":"The TEE set contains 288 image–mask pairs drawn from 71 videos, but the split into 196/40/52 images is described only as ensuring an even distribution across views and cardiac phases. There is no statement that all frames from a given video or patient are confined to a single split. If frames from the same video appear in both the training and test sets, the diffusion model and the downstream nnUNet can memorize video-specific speckle, probe geometry, and anatomy, inflating the reported FID/LPIPS and Dice gains. Please provide a video-exclusive split, ideally at the patient level, and report per-video grouped statistics with confidence intervals. Without this, the central augmentation claim is not supported.","section":"Methods, Data paragraph"},{"comment":"The abstract states that 'adapting only MLP layers suffices for high-fidelity TEE synthesis,' but the Methods state that 'Cross-attention layers are always trained, as they control how the model integrates conditioning signals.' These statements are inconsistent: if cross-attention adapters are always trained, then no configuration actually tests MLP-only adaptation. Please either add an experimental configuration with only MLP LoRA adapters (and no cross-attention training) or revise the claim to say that among the optional adapter groups, adding MLP adapters yields the best trade-off. The exact parameter count of the minimal configuration should also be reported explicitly.","section":"Abstract; Methods, LoRA strategy"},{"comment":"The 'All-Weights' baseline is described as a model trained from scratch on the TEE dataset, yet the results use it to conclude that 'adapters are able to leverage the base model's prior knowledge from TTE data.' A from-scratch baseline cannot isolate the contribution of TTE pretraining; the appropriate control is a fully fine-tuned version of the TTE-pretrained backbone. If 'All-Weights' is actually a fully fine-tuned pretrained model, the text and the Figure 2 caption should say so explicitly. Please add this control or soften the attribution to TTE knowledge.","section":"Methods, Evaluation; Results & Discussion"},{"comment":"The text repeatedly refers to 'Table 1' and 'Figure 2' for FID, LPIPS, SSIM, Dice, and parameter counts, but neither the table nor the figure appears in the manuscript text provided for review. Without the actual numbers, the central quantitative claims cannot be verified. Please ensure that the table and figure are included in the submission and are legible, or state the key numbers directly in the text.","section":"Results & Discussion, Table 1 and Figure 2"}],"minor_comments":[{"comment":"The abstract says 'adapters as small as 10^5 parameters' while the Results say 'around 510,000 parameters'; please use consistent notation and specify the exact minimal parameter count for the best configuration.","section":"Abstract; Results & Discussion"},{"comment":"The FID, LPIPS, and SSIM values and the claimed weak correlation with segmentation impact are reported without confidence intervals; with only 52 test images, FID is high-variance, so consider reporting bootstrap intervals or multiple sampling draws, or explicitly label the correlation analysis as exploratory.","section":"Results & Discussion, image-quality metrics"},{"comment":"Equation (2) uses an unusual underbrace notation around the network output; please rewrite the loss to be fully explicit so that the reader can follow the EDM preconditioning without ambiguity.","section":"Methods, Diffusion formulation"},{"comment":"Reference [4] for LoRA lacks a venue and year, and reference [6] is cited only as an arXiv preprint; please update these entries to their published versions where available.","section":"References"},{"comment":"The term 'All Weights' is used for a model said to be trained from scratch, but the main text refers to it as a baseline for LoRA adaptation; using distinct names such as 'FromScratch' and 'FullFT' would reduce ambiguity.","section":"Figure 2 caption"},{"comment":"The internal TEE dataset is not described as available, which limits reproducibility; adding a data availability statement or a plan for releasing de-identified data would be helpful.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the potential video-level data leakage, because the segmentation gains are the paper's headline application. If the authors cannot implement a video-exclusive split, they should at minimum provide a grouped per-video analysis; otherwise the results should not be presented as generalizable. The abstract/methods contradiction about MLP-only adaptation should also be resolved before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does something useful and mostly honest, but the abstract overstates one finding and the data split is under-described in a way that matters for the headline result.\n\nThe concrete contribution is a pipeline: take a TTE-trained mask-conditioned diffusion model (EDM/UNet), freeze it, attach LoRA adapters to selected layer groups, add a MaskR2 remapping layer to translate TEE masks into the base model's class channels, and fine-tune on 196 real TEE image-mask pairs. The authors then show that synthetic TEE images from this adapted model, mixed 1:1 with real frames, improve nnUNet multiclass Dice over real frames alone, with the largest gains on right-heart structures. That is a practical result for a clinically important but data-scarce modality. MaskR2 is simple and well-motivated, and the paper is right to evaluate with downstream segmentation rather than only FID/LPIPS, whose limits on ultrasound they acknowledge.\n\nNow the soft spots. The abstract's claim that 'adapting only MLP layers suffices' runs head-on into the Methods sentence that cross-attention layers are always trained. If every adapter configuration includes cross-attention LoRA, the MLP-only result isn't actually MLP-only. This matters because the targeted adaptation analysis is one of the three listed contributions. The authors need to either correct the abstract or show a configuration that excludes cross-attention.\n\nSecond, the split. The 288 frames come from 71 videos, and the split is described only as balanced across views and phases. There is no statement that all frames from a video stay in one split. If training and test draw from the same videos, the segmentation gain could be inflated by memorized patient-specific speckle or anatomy. The paper also gives no error bars on the Dice numbers. This is fixable with a per-video grouped split and confidence intervals, but it's load-bearing for the main claim.\n\nMinor: no code/data release (SSM masks are public, but the TEE data and model aren't), and the FID/LPIPS caveat is handled honestly.\n\nWho's this for: people working on synthetic ultrasound generation, cross-modality adaptation, or label-efficient fine-tuning of diffusion models. It deserves a real referee, because the idea is solid and the results are plausible, but the current version needs the split clarified and the internal inconsistency resolved.\n\nRecommendation: send to peer review, with a request for revision on those two points.","headline":"Useful practical adaptation of a TTE diffusion model to TEE with small LoRA adapters and mask remapping, but the abstract's 'MLP-only' claim conflicts with the methods and the train/test split may leak video identities.","tokens_in":8346,"tokens_out":3101,"would_cite":false,"duration_ms":29240,"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 mask-conditioned diffusion model pretrained on transthoracic echo can be adapted to transesophageal echo with low-rank adapters and only 196 real TEE image-mask pairs, generating synthetic frames that improve right-heart segmentation…","keywords":["low-rank adaptation","diffusion models","transesophageal echocardiography","cross-modality generation","mask conditioning","synthetic data augmentation","cardiac segmentation","right-heart structures"],"falsifier":"Run the same LoRA and MaskR$^2$ adaptation protocol but initialize the frozen backbone with random weights instead of TTE-pretrained weights; if the random-backbone version still produces synthetic TEE frames that yield comparable segmentation Dice gains, then transfer from TTE is not what carries the result.","tokens_in":7399,"feed_emoji":"🫀","tokens_out":10250,"duration_ms":96388,"temperature":0.7,"pith_summary":"Deep diffusion models generate realistic images but normally need large training sets, and transesophageal echocardiography (TEE) has very little public data. The paper claims that a diffusion model pretrained on the common transthoracic view (TTE) can be moved to TEE using low-rank adapters of about $10^5$ to $5\\times10^5$ parameters plus a mask-remapping layer, with only 196 real TEE image-mask pairs for training. The adapted model produces synthetic TEE frames that, when mixed with real frames, improve multiclass segmentation Dice, particularly for right-heart structures. The practical stake is a cheap route to controlled synthetic data in a modality where deep learning research is currently scarce.","feed_headline":"196 frames adapt a diffusion model from one echo view to another","feed_subtitle":"Synthetic transesophageal frames made this way boost right-heart segmentation when mixed into real training data.","key_machinery":"The machinery is a frozen, mask-conditioned Elucidated Diffusion Model (EDM) pretrained on TTE, plus two lightweight additions: LoRA adapters, which factor each weight update into low-rank products $BA$ and leave the base weights frozen, and MaskR$^2$, a remapping layer that converts a new dataset's semantic labels into the base model's one-hot conditioning channels using Identity, Reduce, and Repurpose operations. A targeted layer-group study attaches adapters independently to cross-attention, self-attention, convolution, linear, and other layers to show which carry the adaptation. This combination is what lets a model trained on left-heart TTE labels generate TEE images conditioned on masks that include right-heart structures.","core_discovery":"The claim, on the paper's own terms, is that a frozen EDM diffusion backbone pretrained on TTE semantic-mask pairs can be adapted to TEE with LoRA adapters attached to selected layer groups and a MaskR$^2$ layer that maps TEE's mask labels (including right-heart structures) into the TTE model's conditioning channels. With 196 training TEE pairs and roughly $10^5$ to $5\\times10^5$ trainable parameters, the adapted model generates anatomically faithful synthetic TEE images; the paper further claims that adapting the linear/MLP layers, with cross-attention adapters always present, is sufficient for high-fidelity synthesis. When the synthetic frames are mixed 1:1 with real TEE frames to train a segmentation network, global and class-weighted Dice improve over training on real frames alone, particularly for underrepresented right-heart structures. The paper also reports that image-quality scores such as FID, LPIPS, and SSIM correlate only weakly with this downstream segmentation benefit.","pith_inferences":["The Evaluation section itself cautions that FID, LPIPS, and SSIM are built for natural scenes, so the reported image-quality scores should not be read as a complete measure of echo realism; the downstream segmentation gain is the more decision-relevant number.","The abstract's statement that 'adapting only MLP layers suffices' should be read alongside the Methods section's statement that cross-attention adapters are always trained; an ablation that removes cross-attention adaptation entirely would show where the adaptation capacity actually lives.","MaskR$^2$'s Identity/Reduce/Repurpose recipe is a general mechanism for moving any mask-conditioned generator to a dataset with a different label set, so the same pipeline could be tested for other ultrasound views or other imaging modalities whenever a suitable pretrained backbone exists.","The paper tests one training-set size (196 frames); an explicit curve showing how segmentation gain changes as the number of real TEE frames shrinks would map the data-efficiency limit of the approach."],"forward_implications":["With a frozen TTE-pretrained diffusion model and about $10^5$ to $5\\times10^5$ trainable parameters, a research group can generate synthetic TEE frames after collecting only 196 annotated TEE image-mask pairs.","MaskR$^2$ lets the adapted model accept masks with new anatomical structures, such as right-heart chambers, by remapping them into the base model's existing conditioning channels rather than retraining the architecture.","Mixing synthetic TEE frames with real frames in a 1:1 ratio improves multiclass segmentation Dice, with the largest gains on the right-heart classes that are underrepresented in the real training data.","The targeted layer study implies that most adaptation capacity is concentrated in a few layer groups, so the adapters can be merged back into the base model after training and add no extra inference cost.","Because conditioning masks can be drawn from publicly available statistical shape models, the augmentation pipeline does not require a large private TEE dataset to reproduce."],"supporting_citations":[{"why":"Supplies the diffusion framework used as the frozen base generator pretrained on TTE.","marker":"[6]"},{"why":"Provides the public transthoracic echo dataset on which the base model is pretrained.","marker":"[8]"},{"why":"Supplies the LoRA low-rank adapter method used for the targeted parameter-efficient adaptation.","marker":"[4]"},{"why":"Provides the prior TEE generation approach and the plane-extraction pipeline used to obtain statistical-shape-model masks.","marker":"[11]"},{"why":"Provides the statistical shape models of the heart from which out-of-distribution TEE masks are sampled.","marker":"[17]"},{"why":"Supplies the downstream segmentation model trained on real and augmented data and evaluated with Dice metrics.","marker":"[5]"}],"fun_headline_variants":["LoRA adapts TTE diffusion to TEE with just 196 frames","196 frames: LoRA diffusion from TTE to TEE","Tiny LoRA adapters bridge echo views for synthesis","Synthetic TEE from TTE diffusion boosts right-heart Dice","MaskR² + LoRA: 196 frames to generate TEE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the features learned from transthoracic echo transfer well enough to transesophageal echo that a few adapter layers trained on 196 TEE pairs capture the TEE-specific appearance (probe geometry, speckle texture, near-field artifacts) needed to produce realistic, useful synthetic images.","fun_headline_variants_meta":{"raw":{"variants":["LoRA adapts TTE diffusion to TEE with just 196 frames","196 frames: LoRA diffusion from TTE to TEE","Tiny LoRA adapters bridge echo views for synthesis","Synthetic TEE from TTE diffusion boosts right-heart Dice","MaskR² + LoRA: 196 frames to generate TEE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1426,"prompt_tokens":1015,"completion_tokens":411,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":320}},"tokens_in":631,"tokens_out":411,"duration_ms":3998,"temperature":1.0,"reasoning_tokens":320,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:15:21.825704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same LoRA and MaskR$^2$ adaptation protocol but initialize the frozen backbone with random weights instead of TTE-pretrained weights; if the random-backbone version still produces synthetic TEE frames that yield comparable segmentation Dice gains, then transfer from TTE is not what carries the result.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LoRA low-rank adapter method used for the targeted parameter-efficient adaptation."},{"cited_title":"Lecture Notes in Computer Sci- ence (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 14379 LNCS , 43–52 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the prior TEE generation approach and the plane-extraction pipeline used to obtain statistical-shape-model masks."},{"cited_title":"Nature Methods 18, 203–211 (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the downstream segmentation model trained on real and augmented data and evaluated with Dice metrics."}],"review_version":2}