{"id":"19202036-587c-4715-abed-3b13bc5f3d2c","arxiv_id":"2506.12186","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MRI-CORE, a DINOv2-based MRI foundation model initialized from SAM and trained on 6.9 million slices, outperforms SAM, MedSAM, and nnU-Net in few-shot segmentation across most of 13 tasks.","lead":"Researchers trained MRI-CORE, a foundation model on 6.9 million MRI slices from 110,000 scans, and show it improves few-shot segmentation and classification across many body regions. It offers a publicly available starting point for building medical AI with very little labeled data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Epoch-4 checkpoint selection rests on only two in-house datasets; the main few-shot gains survive excluding those tasks, but the generality of the short-pretraining strategy remains conditional.","rationale":"The paper's central claim that MRI-CORE provides large few-shot segmentation gains over SAM, MedSAM, and nnU-Net is supported by detailed per-task tables (Appendix A), consistent positive directions on 11 of 13 tasks, and ablations that rule out several competing explanations (SAM initialization, MAE, sampling strategy, normalization). The reader's weakest-assumption identification is the same as mine: the epoch-4 checkpoint is chosen using only two in-house datasets, and the paper generalizes the short-pretraining conclusion to all tasks. I checked whether this selection could be inflating the headline average by recomputing the mean 3D DSC excluding Private-Bone and Private-Breast: the gain over SAM drops only from 4.38 to 3.67 points, so the main few-shot claim is not an artifact of checkpoint selection. Thus the concern does not refute the central claim, but it does leave a genuine condition on the training-strategy recommendation: no evidence shows that epoch 4 is near-optimal for abdominal, cardiac, spine, or tumor tasks. The paper also has secondary gaps (zero-shot best-k selection, post-hoc FRD analysis, no TRIAD comparison) that the reader noted, but these are not load-bearing for the few-shot result. The verdict should remain CONDITIONAL: the central few-shot claim is credible and reproducible in principle, but the universality of the short-pretraining recipe and the robustness of the zero-shot protocol need additional evidence before the broadest claims are taken at face value.","tokens_in":15575,"tokens_out":16637,"duration_ms":183616,"concrete_test":"Obtain or retrain the intermediate checkpoints at epochs 2, 4, 8, 16, and 32, and run the same few-shot segmentation protocol on a held-out set of tasks not used in any model selection (e.g., AMOS-Liver, AMOS-Pancreas, SPIDER, QIN & ISBI-MR, HVSMR-Vessel, LiverHccSeg-Tumor). Compare average 3D DSC. If a non-epoch-4 checkpoint yields a significantly higher average (e.g., >2 points), the universality of the short-pretraining claim fails and the reported results are checkpoint-dependent; if epoch 4 remains best or within noise, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The checkpoint used for all downstream evaluations (epoch 4) was selected by validation on only Private-Bone and Private-Breast (Sec. 2.1, Fig. 2c). Since these two datasets are also part of the 13-task evaluation, the model-selection procedure could in principle inflate the reported average. However, recomputing the mean 3D DSC over the 11 tasks not used in selection gives MRI-CORE 63.57 vs SAM 59.90 (a 3.67-point gap), compared with a 4.38-point gap over all 13 tasks, so the improvement over SAM persists. The load-bearing unresolved issue is the converse: the paper claims that longer pretraining hurts generalizability, but this is demonstrated only on two musculoskeletal/breast tasks. If a later checkpoint (e.g., epoch 8 or 16) is substantially better on abdominal, cardiac, or spine tasks, the short-pretraining recipe is not universal, and the released epoch-4 model may not be the best representative of the proposed strategy. This is a condition on the training-strategy contribution, not on the existence of a gain over baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces MRI-CORE, a 2D vision foundation model for MRI trained with DINOv2-style self-supervised learning on Duke-110K (6.9 million slices from 116,806 volumes, single institution, no brain MRI). The model is initialized from SAM's ViT-B image encoder and trained for a deliberately short schedule (4 epochs). The authors evaluate few-shot segmentation with five annotated slices on 13 tasks across 10 datasets against SAM, MedSAM, nnU-Net, and a no-pretraining baseline; linear-probe classification on six tasks; and zero-shot segmentation via k-means clustering of patch embeddings. They also present an analysis relating the few-shot gain over SAM to the Fréchet Radiomic Distance between pre-training and downstream data. The central claims are that MRI-CORE outperforms the baselines on average by substantial margins, that short training from SAM initialization is preferable to longer training, and that transfer gains are predicted by FRD.","tokens_in":15729,"tokens_out":5808,"duration_ms":176863,"significance":"If the results hold, MRI-CORE is a valuable community resource: the model weights and code are released under a permissive license, the pre-training dataset is large and spans many body locations and sequence types, and the few-shot segmentation protocol is detailed and evaluated on multiple external public datasets. The paper also provides a useful ablation of training decisions (epochs, learning rate, normalization, initialization, and DINOv2 vs MAE) and a falsifiable, if exploratory, transfer-similarity hypothesis. The main quantitative claims are benchmarked against external models rather than only against the authors' own methods, which is a clear strength. However, the force of several claims is tempered by checkpoint selection on two in-house datasets, an oracle-style zero-shot evaluation, and a correlational FRD analysis with a small number of aggregated data points.","major_comments":[{"comment":"The choice of the epoch 4 checkpoint for all downstream evaluations is made on Private-Bone and Private-Breast only, and these two datasets are also part of the 13-task segmentation benchmark. While the authors state that extended training 'reduced generalizability,' this conclusion is not tested on abdominal, cardiac, or spine tasks; if a later checkpoint is substantially better there, the released model would not be the best representative of the proposed short-pretraining strategy. Recomputing the average over the 11 tasks not used for checkpoint selection still leaves MRI-CORE ahead of SAM (about 3.7 vs 4.4 percentage points in 3D DSC), so the existence of a gain appears to survive, but the general claim about optimal training duration needs either an evaluation of epochs 2, 8, 16, 32, and 64 on the external tasks or a clearly conditional wording.","section":"Sec. 2.1, Sec. 4.2.2"},{"comment":"The quantitative zero-shot segmentation evaluation selects, for each image, the k-means cluster with the highest overlap with the ground-truth object ('find the cluster that overlaps with the target the most'), and the number of clusters k is chosen per dataset as the value giving the best DSC (32 for Private-Bone, 64 for AMOS-Liver). This is an oracle upper bound, not a zero-shot segmentation result: without labels one cannot know which cluster corresponds to the target, and per-dataset selection of k uses test information. The qualitative demonstrations remain interesting, but the quantitative claim of 'leading performance' in zero-shot segmentation is not established by the current protocol.","section":"Sec. 2.4, Sec. 4.3.3, Fig. 5b"},{"comment":"The FRD transfer analysis is based on 10 data points after aggregation, two of which (AMOS and HVSMR) combine multiple tasks and are therefore not independent; the p-values are uncorrected for multiple testing, and no confidence intervals or cross-validation are given. The sentence 'This directly establishes that the more similar a FSL fine-tuning set or test set is to the model's pre-training set, the more that the segmentation performance on the test set will improve' overstates what a correlational, post-hoc analysis can show. Please reframe the conclusion as an exploratory association, report the effective independent sample size or a bootstrap interval, and ideally validate the FRD-based prediction on held-out tasks.","section":"Sec. 2.2.2, Fig. 4"},{"comment":"The paper reports mean and standard deviation for each method and dataset but does not state the number of runs, the source of variance, or whether the same few-shot training slices and seeds were used across methods. With only five training slices per task, margins such as the 1.2-point 3D DSC difference over nnU-Net on Private-Breast or the -1.3-point deficit on SPIDER may be within run-to-run variability. Please add the number of repeated runs, seed sensitivity, and paired significance tests across datasets/tasks (or across repeated draws) to support the claim of a 'significant margin' over the baselines.","section":"Sec. 2.2.1, Tables A1-A13"}],"minor_comments":[{"comment":"The third affiliation is numbered 3 twice; the Department of Computer Science entry should be numbered 4.","section":"Sec. 1 (author affiliations)"},{"comment":"The phrase 'samling with replacement' is a typo for 'sampling with replacement.'","section":"Sec. 2.1"},{"comment":"The sentence 'Qualitatively, as shown in Fig. 2c' should refer to Fig. 2d, since Fig. 2c shows epoch-wise performance and Fig. 2d shows qualitative segmentation examples.","section":"Sec. 2.2.1"},{"comment":"The phrase 'few-show learning' should be 'few-shot learning' in both panels of the caption.","section":"Fig. 4 caption"},{"comment":"The word 'segmetnation' is a typo for 'segmentation.'","section":"Sec. 4.4 (LiverHccSeg)"},{"comment":"It is not stated explicitly whether SAM and MedSAM are fine-tuned with the same adapter protocol as MRI-CORE; please clarify this so that the reader can verify that the comparison is apples-to-apples.","section":"Sec. 4.3.1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: MRI-CORE is a legitimate advance. A DINOv2-style ViT initialized from SAM and fine-tuned for four epochs on 6.9M MRI slices beats SAM, MedSAM, and nnU-Net on most of 13 few-shot segmentation tasks, and the gains are large enough to matter. The model and code are public. The ablations are genuinely informative: sampling without replacement, low LR, slice-wise normalization, and SAM init for both encoders each buy something. I believe the main few-shot result.\n\nThe soft spots are real but not fatal. The checkpoint choice (epoch 4) is based on two in-house musculoskeletal/breast datasets, and the paper uses that same pair in the 13-task average. I recomputed the mean over the other 11 tasks: MRI-CORE still leads SAM by about 3.7 DSC points, so the headline gain does not depend on the selection datasets. What is less supported is the broader claim that longer pretraining actively hurts: that is shown only on those two datasets. If a later epoch is better on abdominal or cardiac tasks, the recipe is less universal than advertised, and the released model may not be the best representative of the strategy. This is a condition on the training-strategy contribution, not on the existence of a gain over baselines.\n\nThe FRD analysis is post-hoc, uses the authors' own metric, and covers 10 points. The negative correlation is consistent with their earlier work, but I would not treat it as a predictive law without a held-out validation. The zero-shot segmentation evaluation is the weakest part: it picks k per dataset to maximize the proxy, so calling it 'zero-shot' is generous; the qualitative figures carry most of the weight. There is also no comparison with TRIAD, the closest 3D MRI foundation model, and the paper explicitly says it did not ablate dataset scale or diversity even though that is a central motivation. Those are addressable gaps, not contradictions.\n\nWho benefits: anyone building MRI-specific segmentation or classification pipelines with limited labels, and groups designing medical SSL recipes. The paper deserves a serious peer review. It is not a desk reject. I would engage with it, but I would ask for a fixed-k zero-shot protocol, a held-out test of the FRD correlation, and at least one later-epoch checkpoint evaluation on non-musculoskeletal tasks.","headline":"A well-ablated MRI foundation model with real few-shot gains; the short-pretraining claim is conditionally supported, and the zero-shot/FRD analyses need tightening.","tokens_in":16351,"tokens_out":2903,"would_cite":true,"duration_ms":28873,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MRI-CORE, a self-supervised vision model trained on 6.9 million MRI slices, claims to outperform general-purpose segmentation foundation models such as SAM and MedSAM across 13 few-shot segmentation tasks, linear-probe classification, and…","keywords":["magnetic resonance imaging","foundation model","self-supervised learning","DINOv2","few-shot segmentation","zero-shot segmentation","transfer learning","Fréchet Radiomic Distance"],"falsifier":"Train MRI-CORE to checkpoints at epochs 2, 4, 8, 16, 32, and 64, then evaluate every downstream task on each checkpoint. If several tasks peak at an epoch other than 4, or if including brain MRI from a second institution shifts the optimum, the claim that a single short schedule is best would be undercut.","tokens_in":15323,"feed_emoji":"🧲","tokens_out":5783,"duration_ms":64825,"temperature":0.7,"pith_summary":"MRI-CORE is a vision foundation model trained exclusively on MRI slices: more than 6 million slices from over 110,000 volumes covering 18 body locations and nine sequence types. The paper's central claim is that a slice-based, MRI-specific model, initialized with the weights of a general image segmentation model and then self-supervised trained for only a few epochs, produces representations that transfer to data-restricted medical tasks better than general-purpose segmentation foundation models. Across 13 few-shot segmentation tasks the model reports an average 3D Dice of 63.4, versus 59.1 for SAM and 48.1 for MedSAM, together with better frozen-feature classification and more semantically coherent zero-shot clusters. The paper also reports that the size of the gain over its own starting point grows as the downstream images become more similar to the pretraining distribution, measured by a radiomic distance metric. The authors note that the pretraining data comes from a single institution, excludes brain MRI, and uses a base-size vision transformer, so the claims are bounded to that regime.","feed_headline":"MRI-only foundation model beats generalist segmenters","feed_subtitle":"Trained on 6.9 million MRI slices, MRI-CORE wins few-shot segmentation, linear-probe classification, and zero-shot tests.","key_machinery":"The central object is the DINOv2 student-teacher self-distillation objective, in which a student vision transformer is trained to match the embeddings of a teacher vision transformer under different image views, with masked-token reconstruction as an additional loss. MRI-CORE adapts this framework by initializing both student and teacher with the ViT-B image encoder of the Segment Anything Model, using slice-wise normalization, sampling without replacement per epoch, and a low learning rate so that only about four epochs of pretraining are run. The short schedule is load-bearing: the paper's ablations show performance peaks at epoch 4 and declines with longer training, which it interprets as preserving SAM's general visual knowledge while adapting it to MRI. The adapter-based few-shot fine-tuning protocol, with adapters inserted into the first and last two transformer blocks, is the mechanism through which transfer is measured.","core_discovery":"The central claim is that a 2D vision transformer pretrained on a large, body-wide set of MRI slices with the DINOv2 self-supervised objective, starting from a Segment-Anything-initialized encoder and trained for a deliberately short schedule of about four epochs, becomes a strong MRI foundation model. The authors show this by fine-tuning only lightweight adapters in a five-slice few-shot setting on ten datasets covering bone, muscle, breast, organs, and tumor segmentation, and reporting that MRI-CORE outperforms SAM, MedSAM, and nnU-Net on average across 2D DSC, 3D DSC, and 3D NSD. They further show that the model's frozen features support linear classifiers for location, sequence, institution, disease, and severity tasks better than the same features from SAM or MedSAM, and that k-means clusters of its spatial features align more closely with anatomical structures in zero-shot settings. A distinctive secondary claim is that the transfer gain over the un-pretrained starting point tracks a distribution-similarity measure: the larger the Fréchet Radiomic Distance between pretraining and downstream images, the smaller the fine-tuning gain.","pith_inferences":["If the Fréchet Radiomic Distance relationship holds generally, it suggests a selection tool: before committing to a foundation model, compute the radiomic distance between its training data and the target dataset and choose the model with the smallest distance.","The exclusion of brain MRI from pretraining means the model's advantage on brain tasks is untested; a natural extension would be to measure whether adding a modest number of brain volumes closes any gap that remains on brain segmentation.","The same short-pretraining recipe could be tested on other imaging modalities such as CT, ultrasound, or pathology by initializing from a strong natural-image encoder and stopping after a handful of epochs; the paper's ablations predict this will beat longer training, but that extrapolation is not demonstrated here."],"forward_implications":["An MRI-only foundation model can be a stronger starting point for label-efficient MRI segmentation than natural-image or mixed-modality foundation models, so downstream users may need far fewer annotated slices for new tasks.","The success of a four-epoch schedule suggests that when initialization is strong, very short self-supervised adaptation can outperform long pretraining by retaining the inductive biases of the initial weights.","The Fréchet Radiomic Distance correlation gives a practical rule of thumb: tasks whose images lie close to the pretraining distribution are the ones most likely to benefit from MRI-CORE, so this distance could be used to predict where transfer will help.","Frozen MRI-CORE features carry usable signals for clinically relevant classification tasks such as anatomy, sequence, institution, and disease severity, which implies they can support linear-probe models trained with modest labels."],"supporting_citations":[{"why":"Supplies the DINOv2 self-supervised learning objective that MRI-CORE adapts for MRI pretraining.","marker":"Oquab et al. 2023"},{"why":"Provides the Segment Anything Model image encoder whose pretrained weights initialize both student and teacher networks.","marker":"Kirillov et al. 2023"},{"why":"MedSAM is the medical-image segmentation foundation model baseline that MRI-CORE is compared against.","marker":"Ma et al. 2024"},{"why":"Provides the adapter insertion and few-shot fine-tuning protocol used for downstream segmentation evaluation.","marker":"Gu et al. 2024"},{"why":"Introduces the Fréchet Radiomic Distance used for the distribution-similarity analysis.","marker":"Konz et al. 2025"},{"why":"Masked autoencoders serve as the alternative self-supervised objective tested as a baseline in the ablations.","marker":"He et al. 2022"},{"why":"Supports the choice of SAM initialization by demonstrating SAM's zero-shot segmentation capabilities on MRI.","marker":"Mazurowski et al. 2023"}],"fun_headline_variants":["MRI-CORE beats SAM and MedSAM in few-shot segmentation","6.9M MRI slices build a strong few-shot segmenter","Transfer gain tracked by distribution similarity in MRI","Body-wide MRI pretraining beats generalist models with tiny adapters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The epoch-4 checkpoint was chosen using validation on two in-house datasets (bone and breast), and the paper assumes that same training length is near-optimal for all 13 downstream tasks; if the ideal pretraining length varies by anatomy or task, the reported average gains could reflect that selection.","fun_headline_variants_meta":{"raw":{"variants":["MRI-CORE beats SAM and MedSAM in few-shot segmentation","6.9M MRI slices build a strong few-shot segmenter","Transfer gain tracked by distribution similarity in MRI","Body-wide MRI pretraining beats generalist models with tiny adapters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1684,"prompt_tokens":935,"completion_tokens":749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":680}},"tokens_in":551,"tokens_out":749,"duration_ms":9981,"temperature":1.0,"reasoning_tokens":680,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:56:46.511961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MRI-CORE to checkpoints at epochs 2, 4, 8, 16, 32, and 64, then evaluate every downstream task on each checkpoint. If several tasks peak at an epoch other than 4, or if including brain MRI from a second institution shifts the optimum, the claim that a single short schedule is best would be undercut.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision, pp 4015--4026","cited_arxiv_id":null,"evidence_quote":"Provides the Segment Anything Model image encoder whose pretrained weights initialize both student and teacher networks."},{"cited_title":"://arxiv.org/abs/2412.01496, https://arxiv.org/abs/2412.01496 arXiv:2412.01496","cited_arxiv_id":null,"evidence_quote":"Introduces the Fréchet Radiomic Distance used for the distribution-similarity analysis."},{"cited_title":"Medical Image Analysis 89:102918","cited_arxiv_id":null,"evidence_quote":"Supports the choice of SAM initialization by demonstrating SAM's zero-shot segmentation capabilities on MRI."}],"review_version":1}