{"id":"8f99a50a-1ff3-493b-ab32-be97e451ce72","arxiv_id":"2505.22682","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Fine-tuning Stable Diffusion with MRI-text pairs yields plausible brain MRI images by field strength and modality, and synthetic images appear to improve a small MRI contrast classification task.","lead":"This paper fine-tunes Stable Diffusion on brain MRI data so that text prompts such as '0.3T brain MRI, T2 contrast' can generate synthetic T1, T2, and FLAIR images. It reports that adding those synthetic images to a small real training set lifts MRI contrast classification from 71% to 96% accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 96.15% classification gain rests on a single-split, single-run comparison that confounds synthetic content with doubled sample size, and §4.3.3 shows weak text-prompt control; the augmentation claim is not yet pinned down.","rationale":"The reader's weakest assumption identifies the classification experiment's transfer and stability problems, and my own reading converges on the same spot. The generation-quality portion of the paper has independent support that I do not see a concrete reason to reject: FID drops sharply after fine-tuning (IncepV3-FID from 317.35 to as low as 88.18, CLIP-FID from 39.94 to 9.98), the qualitative figures show modality-consistent structure, and the datasets used for training (M4Raw and fastMRI) are public. The classification experiment, however, is the only evidence for the 'useful tool' part of the central claim, and it is not designed to isolate the contribution of synthetic data. The training-set sizes are tiny, the validation set is only about eight images per model after the 8:2 split, and all validation accuracies are reported as 100%, which §5.4.1 itself flags as a likely overfitting signal. A single run at 100 epochs with no reported seed or checkpoint selection makes the 71.15% versus 96.15% comparison statistically unanchored. Moreover, because the augmented condition doubles the number of training images, the observed gain could be a sample-size effect rather than evidence that the synthetic images carry label-relevant contrast information that transfers across scanner and population. The manuscript's own discussion in §6.2 lists the failure to fine-tune the text encoder as a limitation, and §4.3.3 shows that prompt modifications do not visibly change generation, so the 'text-prompted' characterization is also weaker than the title and abstract suggest. These are not internal inconsistencies in the FID evaluation; they are missing controls and missing repetitions in the one experiment that supports the practical claim. Therefore I agree with the CONDITIONAL verdict: the direction is plausible and worth confirming, but the current evidence does not support the strength of the stated conclusions. No verdict change is needed beyond what the reader already recommended.","tokens_in":23703,"tokens_out":6780,"duration_ms":59125,"concrete_test":"Run a leave-one-subject-out cross-validation on the five 0.35T subjects (train on four, test on one), repeated over at least five random seeds, for four conditions: (a) 0.35T real only, (b) 0.35T real + synthetic 0.3T images, (c) 0.35T real + 3T real images with the same added count, and (d) 0.35T real with doubled on-the-fly augmentation. Report mean ± standard deviation of accuracy, precision, recall, and F1, plus McNemar's test between paired predictions of conditions (a) and (b). If (b) is not significantly better than (c) or (d), the claim that SD-generated images specifically improve scarce-sample classification is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that text-prompted MRI generation is a useful tool for medical AI depends on the classification experiment in §5. The key comparison is between 13 real 0.35T images per class (71.15% test accuracy) and the same 13 real images plus 13 synthetic 0.3T images per class (96.15%). This design confounds the effect of synthetic data with the effect of simply having more training samples. The control condition, 3T real data (18 images per class), is not an adequate control because it changes the domain as well as the sample count; no condition adds an equivalent amount of non-synthetic data to the 0.35T training set. The manuscript itself warns in §5.4.1 that 100% validation accuracy on the small validation split may indicate overfitting, and §5.4.2 acknowledges that the test set comes from a different scanner brand, field strength, and patient population than the synthetic training set. With one subject-level split, no repeated seeds, and no early-stopping rule, the 25-percentage-point jump cannot be separated from sampling variability, initialization luck, or the doubled training-set size. Additionally, §4.3.3 reports that modifying the text prompt with imaging requirements or unseen features has little effect on generated images, so the text-prompted component of the pipeline is not demonstrated to be the controlled variable. The FID/MS-SSIM improvements are credible and are not the main problem; the downstream utility claim is the load-bearing and under-supported part.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper fine-tunes Stable Diffusion v1.5 on brain MRI data from the 0.3T M4Raw dataset and the 3T fastMRI dataset, using text prompts that specify field strength, modality, and slice position under two fine-tuning strategies (DreamBooth and UNet fine-tuning). It evaluates generation quality with Inception-V3 FID, CLIP-based FID, and MS-SSIM, and then tests downstream utility by training a ResNet50 classifier on a small 0.35T clinical dataset, with and without synthetic 0.3T images added to the training set. The authors report that fine-tuning substantially improves FID over the original model and that augmenting the 0.35T training set with synthetic images raises contrast-classification accuracy from 71.15% to 96.15%, concluding that text-prompted MRI generation is feasible and can serve as a useful tool for medical AI applications.","tokens_in":23934,"tokens_out":5417,"duration_ms":40113,"significance":"If the downstream claim were established, the paper would provide practical evidence that synthetic MRI can augment scarce clinical training sets, which is a relevant and timely direction. The experiment's use of an external 0.35T dataset from a different scanner and patient population than the generative training data is a genuine strength, as is the evaluation of generation quality with two complementary FID variants. However, the classification experiment that carries the central claim confounds synthetic data with doubled sample size and lacks any statistical controls, and Section 4.3.3 shows that the text-prompt component has little observable effect. The generation-quality improvement itself is credible and could be of interest, but it is of incremental value without the downstream validation.","major_comments":[{"comment":"The key comparison that supports the paper's central claim confounds synthetic content with doubled sample size: the condition '0.35T Real + Synthetic Data' has 26 images per class while the '0.35T Real Data' condition has 13, and the only other larger condition ('3T Real Data', 18 per class) also changes the domain. There is no control condition that adds an equivalent number of real 0.35T images to the 13-image training set, so the accuracy gain from 71.15% to 96.15% cannot be attributed specifically to the synthetic images. This is load-bearing for the conclusion that generated images can augment scarce clinical training sets.","section":"§5.2, Table 5-1; §5.4.2, Table 5-2"},{"comment":"The classification experiment is a single run on a single subject-level split, with one training subject, four test subjects, no repeated seeds, no cross-validation, and no statistical significance test. The validation split is only 20% of a very small training set, and §5.4.1 itself notes that 100% validation accuracy may indicate overfitting. With 13–26 training images per class and a 156-image test set, the 25-percentage-point difference between the best and worst conditions could plausibly arise from sampling variability or initialization luck; the manuscript should report multiple splits and seeds with confidence intervals or an appropriate statistical test.","section":"§5.3, §5.4.1, §5.4.2"},{"comment":"The manuscript reports that modifying the text prompt with imaging requirements or with features not present during training has little effect on the generated images, and that the same result holds for the UNet fine-tuned models. This directly weakens the 'text-prompted' element that is central to the paper's title and abstract: the observed improvements after fine-tuning could come from learning the domain distribution with fixed class labels rather than from semantic text control. The authors should either demonstrate measurable text-prompt control (for example, through a quantitative text-image alignment metric) or reframe the claims as class-conditional rather than text-prompted generation.","section":"§4.3.3, Figures 4-11 to 4-13"}],"minor_comments":[{"comment":"The heading '4. Experimental Results and and Discussion' contains a duplicated 'and'; the manuscript would benefit from a thorough editorial pass, as many sections read like a raw machine translation.","section":"Chapter 4 heading"},{"comment":"It is not specified how the aggregate FID and MS-SSIM values in Table 4-2 are computed (for example, pooled features over all modalities and field strengths, or averaged per-class values), which makes it difficult to compare them with the per-modality tables that follow.","section":"§3.3, Tables 4-2 to 4-4"},{"comment":"No information is given about the number of generated images used for the FID and MS-SSIM computations, or about the random seeds used for sampling; reporting these details would substantially improve reproducibility.","section":"§4.4"},{"comment":"The synthetic data for the classification experiment are generated with random slice numbers between 1 and 18, but the M4Raw-derived training distribution contains anatomy that varies across slice positions; the paper does not discuss whether this random slice assignment could affect the label-relevant contrast information.","section":"§5.2, Table 5-1"},{"comment":"The choice to retain only the first 10 fastMRI slices and all 18 M4Raw slices is described as filtering unusable scan layers, but no quantitative evidence is provided that this threshold captures a representative slice range; a figure or histogram of selected versus discarded slices would be helpful.","section":"§3.1.2"},{"comment":"Section 6.2 acknowledges that only the UNet was fine-tuned and the text encoder was left frozen; given the weak prompt sensitivity documented in §4.3.3, the discussion could more explicitly connect this architectural choice to the prompt-control result.","section":"§6.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a translated thesis chapter and would benefit from professional English editing. The main issue for the editor is that the headline classification gain is not yet supported because of the confound in the augmentation experiment and the lack of repeated runs; I would not recommend acceptance until this is addressed with a proper control condition that adds an equivalent number of real images, and until the text-prompt control limitation is either resolved or the claims are reframed accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nWhat you should know about this paper: it's a thesis-style application of Stable Diffusion fine-tuning to MRI generation, and the headline number (96.15% classification accuracy) is not as solid as it looks. But the underlying generation story is more believable than the abstract suggests, and the authors are honest about some limitations.\n\nThe new thing is the combination: they fine-tune SD v1.5 on text-image pairs from M4Raw (0.3T) and fastMRI (3T), with prompts encoding field strength, modality, and slice index, and they test synthetic images on an external 0.35T dataset. That external test is a genuinely useful check because the synthetic and test data come from different scanners and populations. The FID improvements after fine-tuning are large (IncepV3 from 317 to 88 in the best UNet config, CLIP from ~40 to ~10) and consistent across modalities. The MS-SSIM numbers show the expected diversity/quality tradeoff. I consider those generation results credible.\n\nThe soft spot is the downstream claim. The key comparison is 13 real images per class vs. 13 real + 13 synthetic. That confounds the value of synthetic data with the value of having twice as many training samples. The 3T condition uses 18 real images per class but from a different domain, so it doesn't control for sample size either. Everything is single-run, with no seeds or intervals, and validation accuracy hits 100%; the authors themselves flag the overfitting risk in §5.4.1. On top of that, §4.3.3 shows that editing prompts to add imaging requirements or unseen features has little effect on generated images, so the \"text-prompted\" part of the claim is not actually the demonstrated controlled variable. The paper would be much stronger with a same-domain real-data augmentation condition, several seeds per condition, and ideally a comparison of text-controlled vs. class-conditional generation if prompt control matters.\n\nThere is no code or weights, and the machine-translated text has rough spots, but the experimental tables are coherent enough to follow. This paper is for readers interested in low-field MRI synthesis and data augmentation for small medical datasets. I would send it to a serious referee: the evaluation needs major revision, but the direction is legitimate and the FID findings are worth checking. I wouldn't cite it in its current form.\n\nBest,\n[Your name]","headline":"The generation results are plausible, but the headline augmentation claim confounds synthetic data with doubled sample size and lacks repeated runs, so the 96.15% figure is not yet pinned down.","tokens_in":24526,"tokens_out":2331,"would_cite":false,"duration_ms":28037,"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":"This paper claims that text-prompted MRI generation is feasible and practically useful: fine-tuning Stable Diffusion on paired MRI-text data yields T1, T2, and FLAIR brain images, and synthetic images can raise contrast-classification…","keywords":["Magnetic Resonance Images","Image Generation","Diffusion Model","Text Guidance","Stable Diffusion","Data Augmentation","MRI Contrast Classification","Low-Field MRI"],"falsifier":"A reader could run the same four training-set conditions on a larger, multi-scanner test set (say 300+ slices from several 0.3–0.35T devices with mixed patient ages). If the accuracy gain from adding synthetic images does not reproduce, or if synthetic-only training falls to chance, the paper's central claim is refuted.","tokens_in":23391,"feed_emoji":"🧠","tokens_out":8547,"duration_ms":55904,"temperature":0.7,"pith_summary":"MRI datasets are costly, scarce, and privacy-restricted, which starves medical AI of training data. The paper argues that text-prompted image generation can help: a Stable Diffusion model fine-tuned on paired MRI-text data produces plausible T1, T2, and FLAIR brain slices from short descriptions of field strength, contrast, and slice position. Adding thirteen synthetic slices per class to a small real training set lifts a three-class contrast classifier from 71.15% to 96.15% accuracy in the paper's test, suggesting synthetic data can substitute for scarce real acquisitions. A sympathetic reading is that the contribution is a demonstration of feasibility plus a recipe: dataset construction, two fine-tuning strategies, and two evaluation metrics, rather than a new generative architecture.","feed_headline":"Text-prompted MRI images lift classifier accuracy to 96%","feed_subtitle":"Fine-tuned Stable Diffusion makes synthetic T1/T2/FLAIR brain scans that raise scarce-sample classification from 71% to 96%.","key_machinery":"The mechanism that carries the argument is the cross-attention-conditioned UNet in Stable Diffusion, which injects text-prompt embeddings into the denoising latent process, so the generator learns to associate phrases like \"0.3T brain MRI, slice 5, T1 contrast\" with image content. Two fine-tuning strategies operationalize this: DreamBooth (full-parameter fine-tuning with a rare identifier plus class noun, which requires a single shared prompt per batch) and direct UNet fine-tuning with per-slice structured prompts. FID, computed with both InceptionV3 and CLIP encoders, measures distributional closeness and semantic consistency of the generated set to real images; MS-SSIM measures structural similarity among generated images, hence inversely their diversity. The downstream demonstration freezes ResNet50's convolutional backbone and trains only the fully connected layer, a design choice that limits overfitting in the scarce-sample setting.","core_discovery":"The paper's central claim is that the Stable Diffusion architecture, originally trained on natural images, can be repurposed into a controllable MRI generator by fine-tuning on paired image-text data from the 0.3T M4Raw and 3T fastMRI datasets. Two fine-tuning routes are compared: DreamBooth, which binds generic modality and field-strength concepts to rare identifiers, and direct fine-tuning of the UNet denoiser with structured prompts encoding field strength, slice index, and contrast type. Both routes improve on the unmodified model: IncepV3-FID falls from 317.35 to 88.18 at the best configuration, CLIP-FID from 39.94 to 9.98, and MS-SSIM rises from 0.05 to 0.29–0.38, indicating better fidelity at some cost in diversity. The downstream experiment is the strongest claim: with only one real subject's 13 slices per contrast as training data, adding 13 generated slices per contrast raises a frozen-backbone ResNet50 classifier's test accuracy on a 156-image 0.35T test set from 71.15% to 96.15%, which the authors take as evidence that generated images carry label-relevant contrast structure.","pith_inferences":["A control experiment the paper does not run would distinguish generic regularization from synthetic-data value: training with 13 additional real slices from a matched scanner instead of 13 synthetic slices, and checking whether the accuracy gain persists.","The paper's prompt-variation results (adding unseen features to the prompt barely changes the output) imply the model's text conditioning is shallow; unfreezing the text encoder or training with more diverse prompts would be a testable extension.","Because the generator trains on healthy young volunteers, augmenting datasets of elderly or diseased patients with these synthetic images risks biasing classifiers toward the healthy distribution; a clinical pipeline would need to measure that bias."],"forward_implications":["Adding synthetic low-field MRI slices to a small real training set can raise contrast-classification accuracy from 71.15% to 96.15%, so text-prompted generation is a viable data-augmentation route when real MRI data are scarce.","Synthetic-only training nearly matches real-data training (69.87% vs 71.15% accuracy), meaning generated images retain enough contrast-specific structure to stand in for real images when none are available.","Fine-tuning with structured text prompts yields FID improvements across all six field-strength and contrast combinations (IncepV3-FID down to 70–133 depending on setting), indicating the recipe generalizes across acquisition parameters.","Because the model can be prompted with slice index, it learns slice-position-specific anatomy, enabling dataset expansion that preserves anatomical plausibility at chosen positions."],"supporting_citations":[{"why":"Supplies the latent diffusion model (Stable Diffusion) that the paper fine-tunes for text-prompted MRI generation.","marker":"[18]"},{"why":"Gives the DreamBooth fine-tuning strategy used to bind MRI modality and field-strength concepts to rare identifiers.","marker":"[35]"},{"why":"Provides the 0.3T multi-contrast brain MRI training data (M4Raw) used for fine-tuning and synthetic generation.","marker":"[33]"},{"why":"Provides the 3T brain MRI training data (fastMRI) used to cover high-field-strength generation.","marker":"[34]"},{"why":"Defines the FID metric used to evaluate the fidelity and distributional closeness of generated images.","marker":"[36]"},{"why":"Defines the MS-SSIM metric used to evaluate structural similarity and, inversely, diversity of generated images.","marker":"[37]"}],"fun_headline_variants":["Text-prompted MRI synthesis lifts classifier to 96%","Stable Diffusion fine-tuned to generate brain MRI scans","Synthetic MRI from text boosts rare-sample accuracy","From text to MRI: synthetic scans raise accuracy to 96%","Text-guided MRI model improves classifier from 71% to 96%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that 13 generated slices from a model trained on healthy 0.3T volunteers can transfer contrast-defining structure to a test set from a different scanner, field strength, and patient population (elderly patients), and that a single train-test split with perfect validation accuracy is enough to trust the 96.15% result.","fun_headline_variants_meta":{"raw":{"variants":["Text-prompted MRI synthesis lifts classifier to 96%","Stable Diffusion fine-tuned to generate brain MRI scans","Synthetic MRI from text boosts rare-sample accuracy","From text to MRI: synthetic scans raise accuracy to 96%","Text-guided MRI model improves classifier from 71% to 96%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1302,"prompt_tokens":1007,"completion_tokens":295,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":210}},"tokens_in":623,"tokens_out":295,"duration_ms":4495,"temperature":1.0,"reasoning_tokens":210,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:47:59.043655+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could run the same four training-set conditions on a larger, multi-scanner test set (say 300+ slices from several 0.3–0.35T devices with mixed patient ages). If the accuracy gain from adding synthetic images does not reproduce, or if synthetic-only training falls to chance, the paper's central claim is refuted.","supporting_citations":[{"cited_title":"High -resolution image synthesis with latent diffusion models[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion model (Stable Diffusion) that the paper fine-tunes for text-prompted MRI generation."},{"cited_title":"Dreambooth: Fine tuning text-to-image diffusion models for subject -driven generation[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Gives the DreamBooth fine-tuning strategy used to bind MRI modality and field-strength concepts to rare identifiers."},{"cited_title":"M4Raw: A multi-contrast, multi-repetition, multi-channel MRI k-space dataset for low-field MRI research[J]","cited_arxiv_id":null,"evidence_quote":"Provides the 0.3T multi-contrast brain MRI training data (M4Raw) used for fine-tuning and synthetic generation."},{"cited_title":"GANs trained by a two time -scale update rule converge to a local nash equilibrium[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Defines the FID metric used to evaluate the fidelity and distributional closeness of generated images."},{"cited_title":"Multiscale structural similarity for image quality assessment[C]//The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003: 1398-1402","cited_arxiv_id":null,"evidence_quote":"Defines the MS-SSIM metric used to evaluate structural similarity and, inversely, diversity of generated images."}],"review_version":1}