Pith. sign in

REVIEW 4 major objections 8 minor 28 references

Panda: Unsupervised Pelvic Anomaly Detection for Real-Time MR Imaging

T0 review · 4 major / 8 minor · reviewed 2026-07-31 · grok-4.5

Pith's one-line read A frozen vision transformer learns normal pelvic MRI from healthy scans alone and flags anomalies in real time without labels.

desk verdict Solid domain adaptation of Dinomaly/DINOv3 for pelvic MRI UAD with real clinical framing, but the headline 88% AUROC was selected on the test set and frame-level recall is only ~12%. read the letter →

arxiv 2607.24703 v1 pith:CTEF5IW2 submitted 2026-07-27 cs.CV

classification cs.CV
keywords UnsupervisedAnomalyDetectionPelvicMRIVisionTransformersFoundationModelsReal-timeImagingDINOv3UterineMyoma
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Pelvic MRI is hard to read in real time: motion, deformation, and rare heterogeneous disease make supervised labels impractical. This paper argues that an unsupervised reconstruction setup can learn what healthy pelvic anatomy looks like from routine cases only, then mark deviations as spatial anomaly maps at the scanner. The method freezes a large pretrained vision transformer, forces a noisy bottleneck so the decoder cannot simply copy the input, and scores anomalies by cosine distance between encoder and decoder tokens. On a curated uterine-myoma test set it reports 88% pixel-level AUROC, about 95% frame-level specificity, and 40.5 slices per second. If those numbers hold under real deployment, radiologists could get immediate localized flags and adapt protocols during the exam rather than after the fact.

What carries the argument

Dinomaly-style reconstruction at representation level: a frozen DINOv3 encoder supplies stable normative features; a noisy MLP bottleneck blocks identity mapping; a Linear Attention decoder reconstructs grouped low- and high-level tokens; anomaly maps are per-token cosine distance between encoder and decoder features, reduced to frame scores by healthy-calibrated percentile thresholds.

What would settle it

Retrain and retune only on a held-out healthy validation split, then measure pixel AUROC, frame sensitivity/specificity, and latency on a fresh multi-site pelvic MRI cohort that includes small low-contrast lesions and non-myoma pathologies never seen during tuning; a large drop below the reported 88% AUROC or collapse of useful frame flags would falsify the central performance claim.

Watch

Extended reading notes

Core claim

An adapted Dinomaly architecture—frozen DINOv3 ViT-L/16 encoder, dropout-noisy MLP bottleneck, and Linear Attention decoder—learns normative pelvic T2-weighted MRI from healthy volumes only and localizes anomalies via per-token cosine distance, reaching 88.06% pixel-level AUROC and 95.45% frame-level specificity at 40.5 slices/s on out-of-distribution uterine myoma data, which the authors present as meeting real-time clinical feedback needs.

Load-bearing premise

That choosing hyperparameters by directly maximizing pixel AUROC on the same test set still gives an honest picture of how well the method will work on new patients and scanners.

Editorial extensions

If this is right

  • Scanner-side spatial anomaly maps could support adaptive protocol changes during pelvic MRI without waiting for offline review.
  • Unsupervised healthy-only training removes the need to annotate rare heterogeneous pelvic adverse events for detection.
  • Reconstruction-based and generative detectors show complementary strengths, motivating fused high-specificity plus high-sensitivity systems.
  • Coarse 28×28 token grids remain the main barrier to small-lesion recall, so finer tokens or multi-scale scoring would be the next design lever.
  • The same frozen-encoder recipe is positioned for interventional pelvic MRI where labeled adverse events are scarce.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If healthy-calibrated high-specificity thresholds stay trusted in the reading room, the tool is more likely to be used as a quiet second reader than as a high-recall screening net.
  • Cross-vendor OOD success on myomas does not yet imply the same maps will light up endometriosis, adenomyosis, or instrument artifacts without further healthy diversity.
  • Pairing this detector with a high-sensitivity generative branch could raise small-lesion recall without abandoning the frozen normative reference.
  • Real-time slice flagging at ~40 Hz is fast enough that the bottleneck for clinical impact shifts from compute to how radiologists act on the maps during acquisition.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The manuscript adapts the Dinomaly unsupervised anomaly detection architecture to sagittal T2w pelvic MRI: a frozen DINOv3 ViT-L/16 encoder, a dropout-noised MLP bottleneck, and a Linear Attention decoder trained with a group-wise cosine reconstruction loss on 207 healthy volumes (plus 800 synthetic images). Per-token cosine distance between encoder and decoder features yields pixel-level anomaly maps; a frame-level score based on the spatial extent of supra-threshold pixels is calibrated on healthy training data only. On a curated subset of the Uterine Myoma Dataset (198 anomalous slices, 102 healthy slices) the authors report 88.06% pixel-level AUROC, 95.45% frame-level specificity at 11.61% sensitivity, and 40.5 slices/s inference, positioning the system as a real-time scanner-side decision support tool, with a comparison to a ResVAE baseline and an ablation study.

Significance. If the results hold up under a clean selection protocol, the work is a useful and timely contribution: unsupervised anomaly detection for pelvic MRI is underexplored, the evaluation is on a public dataset (UMD) and framed as a reproducible benchmark, and the system-level properties (frozen encoder, 40.5 slices/s on a mid-range GPU) are directly relevant to scanner-side deployment. Notable strengths: the frame-level thresholds are calibrated exclusively on healthy training data and thus provide a genuinely leakage-free operating point; the authors are transparent about the oracle nature of the pixel threshold; the ablation (Table 2) includes a data-scaling experiment and a ViT-S speed/accuracy trade-off; and the head-to-head with a reconstruction-based ResVAE, showing complementary sensitivity/specificity profiles, is informative for the field. The paper also reports its weak frame-level sensitivity openly rather than hiding it.

major comments (4)
  1. [§3.2] Hyperparameters were selected by Bayesian optimization over 40 configurations directly maximizing pixel-level AUROC on D_test, and the same D_test is then used to report the headline 88.06% AUROC. The defense offered (no gradient updates on test data; validation loss uncorrelated with detection, rho_S=+0.094) does not address selection bias: taking the argmax of 40 noisy performance estimates on the reporting set is adaptive analysis, and with n=198 anomalous / 102 healthy slices the per-config AUROC standard error is on the order of 2-3 points, so the optimistic shift from selection is plausibly of comparable magnitude. This is load-bearing because the abstract's central claim is this number. The fix is feasible within scope: carve a validation split from D_test (or use nested resampling), re-select hyperparameters on it, and report the selected model on the untouched remainder; alterna
  2. [Abstract and §5, with Table 1] The abstract states the framework 'meets real-time clinical deployment requirements', but the only operating point free of test-set tuning — the frame-level score with thresholds calibrated exclusively on healthy D_train slices — achieves 11.61% sensitivity at 95.45% specificity, i.e., roughly 9 of 10 anomalous slices are not flagged. A scanner-side triage tool that misses ~88% of anomalous slices is a very different clinical proposition than the abstract conveys. Speed (40.5 slices/s) is necessary but not sufficient for the deployment claim. The abstract and §5 should be rewritten to state the actual operating characteristics, and the claim of clinical utility should be tied to the pixel-level maps as a localization aid on flagged/visibly suspicious slices rather than to frame-level detection. As written, the abstract cherry-picks the tuned pixel-level number and the high specificity wh
  3. [§3.3 and Table 1] The pixel binarization threshold tau_px is set by maximizing Dice across D_test. The text calls this an 'upper bound', which is honest, but the resulting accuracy 93.62% / sensitivity 38.49% / specificity 95.10% / F1 23.87% are then presented in Table 1 as the paper's primary pixel-level metrics without any marker distinguishing them as oracle (per-dataset optimal-threshold) values. Table 1 should annotate these rows as upper-bound/oracle, and the comparison with ResVAE in §4 should state explicitly whether the ResVAE numbers use the same oracle-threshold protocol (the slightly different values quoted in the text — 87.85 vs. 88.06, 38.39 vs. 38.49, 95.02 vs. 95.10 — suggest the two may not be from identical evaluation runs).
  4. [§3.1 (Evaluation Data)] The evaluation uses one pre-selected slice per examination — the slice containing the annotated lesion — for the 198 anomalous cases, and healthy slices sampled at distance >=2 from lesions. Two concerns: (i) selecting the lesion-containing slice conditions on the annotation and removes the natural prevalence and difficulty distribution of full volumes, inflating both pixel AUROC and frame scores relative to whole-volume operation; (ii) anomalous and healthy slices come from the same UMD exams, so the frame-level task partly tests within-exam slice discrimination rather than detection of anomalous exams. No confidence intervals are reported anywhere despite the small sample. At minimum, report bootstrap CIs for Table 1 and add a whole-volume evaluation (score all slices, aggregate per exam) as a secondary result, or explicitly scope the claims to the curated per-slice benchmark.
minor comments (8)
  1. [Table 2] Table 2 lists '8 Attention Heads' as an ablation, but §3.2 states the full model already uses eight attention heads; presumably the ablation varies head count from the base. Similarly '0 Decoder MLP Ratio' is unclear notation for what is presumably a variant MLP ratio. Please clarify the ablation axes.
  2. [§4, Comparison with ResVAE Baseline] Numerical inconsistencies between §4 text and Table 1 for the ResVAE comparison: text gives 87.85/38.39/95.02 for the proposed method where Table 1 gives 88.06/38.49/95.10. Please reconcile.
  3. [Table 1] Frame-level F1 is reported as '-' in Table 1; either compute it or state why it is not meaningful at this operating point.
  4. [§3.2] The over-precise hyperparameter values (dropout p=0.136, MLP ratio 2.097, lr 4.24e-4) are clearly artifacts of the Bayesian optimization; once the selection protocol is revised per the major comment, rounding or justifying these would improve readability.
  5. [Fig. 3] Fig. 3: the color legend (yellow/green/red) is described only in the caption; consider in-figure labeling. State in the caption whether the four cases were selected randomly or as representative/best cases.
  6. [§4] Only one baseline (ResVAE, from the authors' own prior work) is compared. Classical feature-embedding anomaly detection methods (e.g., PatchCore, PaDiM, or EfficientAD) are natural, strong, training-free-or-light baselines on frozen DINO features and would contextualize how much the decoder/bottleneck machinery adds over k-NN or Mahalanobis scoring on the frozen encoder. This is not blocking but would substantially strengthen the empirical claim.
  7. [§3.1] §3.1 first describes D_test as '198 anomalous volumes' and later '198 anomalous slices' (one per examination); please make the volume/exam/slice counts consistent throughout. Also state how many of the 102 healthy slices come from UMD vs. the 13 held-out low-field volumes.
  8. [§4] Latency (0.0247 s/slice, RTX 3080): report whether this includes preprocessing (resizing to 224x224, percentile scaling) and upsampling of the anomaly map, since scanner-side deployment would include these.

Circularity Check

2 steps flagged · score 4.0 of 10

Headline pixel AUROC is the direct objective of Bayesian HP search on D_test; binary pixel metrics use a Dice-maximizing threshold also fit on D_test (acknowledged as upper bound). Method itself is not definitionally tautological.

  1. fitted input called prediction [§3.2 Architecture and Training; Abstract; Table 1]
    "Hyperparameters were selected via Bayesian optimization over 40 configurations directly maximizing pixel-level AUROC on D_test, with no gradient updates derived from test data, as validation loss showed negligible correlation with detection performance (ρS = +0.094, p = 0.566)."

    The primary reported result (pixel AUROC 88.06%) is exactly the objective that the 40-configuration Bayesian search maximized on the same D_test used for final reporting. No gradient leakage does not remove selection bias: the published number is the max of 40 correlated test estimates on a small set (198 anomalous slices), so the headline metric is optimistically forced by the selection procedure rather than being a single pre-specified out-of-sample evaluation.

  2. fitted input called prediction [§3.3 Evaluation (Pixel-Level Anomaly Scoring); Table 1 binary metrics]
    "For binary metrics, a pixel threshold τpx is determined by maximizing the Dice coefficient across Dtest, providing an upper bound on performance."

    Accuracy 93.62%, sensitivity 38.49%, specificity 95.10%, and F1 23.87% in Table 1 are computed at a threshold chosen by maximizing Dice on the identical test set. By construction these are the best binary operating point on D_test, not a pre-committed threshold. The paper’s “upper bound” language is honest, but the fitted threshold is still presented alongside the main results as if it were an evaluated operating characteristic.

full rationale

This is an empirical methods paper, not a first-principles derivation, so classical self-definitional circularity is absent: the model is trained only on healthy D_train, anomalies are scored by encoder–decoder cosine distance, and frame-level thresholds (τ_ref, τ_slice) are calibrated exclusively on healthy training slices. The independent content—frozen DINOv3 encoder, noisy MLP bottleneck, Linear Attention decoder, real-time latency, and the ResVAE comparison—does not reduce to its inputs by construction. What does qualify as partial fitted-input circularity is the reporting chain for the central pixel-level numbers. Section 3.2 states that hyperparameters were chosen by Bayesian optimization over 40 configurations that directly maximized pixel-level AUROC on D_test; the abstract and Table 1 then lead with that same 88.06% AUROC. Selecting the argmax of 40 noisy estimates on the evaluation set and reporting that value as the primary result is classic selection bias: the reported figure is optimistically shifted by the selection procedure even though no gradient steps touch test labels. Separately, §3.3 sets the binary pixel threshold τ_px by maximizing Dice on D_test and reports accuracy/sensitivity/specificity/F1 at that threshold; the paper correctly labels this an upper bound, so the circularity is disclosed but the numbers still appear in the main results table. Self-citation to the authors’ prior ResVAE work [8] is only a baseline, not a load-bearing uniqueness claim. Overall this is moderate metric-selection circularity (score 4), not a tautological derivation: the method can still succeed or fail, but the headline pixel metrics are not an untainted out-of-sample estimate.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The claim rests on transfer of frozen natural-image DINOv3 features to pelvic MRI, on reconstruction-disagreement as anomaly, on the purity of 'healthy' training data, and on a large set of fitted architectural and threshold parameters—several selected on the test set itself. No new physical entities are postulated; the invented pieces are engineering modules (noisy MLP bottleneck usage, LA decoder stack, scoring rules) taken or lightly modified from Dinomaly-style UAD.

free parameters (6)
  • Bayesian-optimized hyperparameter set (40 configs maximizing pixel AUROC on D_test) = lr=4.24e-4, dropout p=0.136, MLP ratio=2.097, among others
    Learning rate, dropout, MLP ratio, and related knobs were chosen by directly maximizing the reported test metric rather than a held-out validation objective.
  • Hard-token gradient shrinking percentile and down-weight = 33rd percentile, factor 0.1
    Tokens below the 33rd percentile of per-token error are down-weighted by 0.1 from epoch 1; choice affects sensitivity to subtle deviations.
  • τ_ref and τ_slice (frame scoring thresholds) = 95th percentile each
    Set to 95th percentiles of pixel anomaly scores and of s(x) on healthy D_train slices; define the high-specificity operating point.
  • τ_px (pixel binarization threshold) = Dice-maximizing on D_test
    Chosen by maximizing Dice on D_test; directly shapes accuracy/precision/sensitivity/F1 in Table 1.
  • Synthetic near-duplicate cosine filter threshold = 0.95
    Synthetic images excluded if cosine similarity to existing training samples ≥ 0.95.
  • Uterine crop and input sizes = 96×96 crop → 224×224
    96×96×n bounding box resized to 224×224 for ViT; defines the spatial support of all maps.
assumptions (6)
  • domain assumption Frozen DINOv3 intermediate features form a stable normative reference for pelvic T2w MRI without task-specific encoder fine-tuning.
    Load-bearing for the entire pipeline (§2.2, §3.2); justified by ablation robustness but not proven across pathologies.
  • domain assumption Elevated per-token cosine distance between encoder and decoder features indicates clinically relevant anomaly rather than domain shift, motion, or crop edge effects.
    Problem formulation §2.1 and anomaly map definition; bowel FPs in Fig. 3 show the assumption is imperfect.
  • domain assumption Training volumes labeled healthy contain no pathologies, major artifacts, or clinically significant deviations.
    §2.1 definition of D_train; residual unlabeled disease would poison the normative model.
  • ad hoc to paper Linear Attention's globally distributed aggregation is better matched to subtle contextual pelvic inconsistencies than softmax attention.
    Motivated in §2.2; not independently validated against a softmax-decoder control in the ablations.
  • domain assumption Group-to-group cosine reconstruction loss on layers 3–6 and 7–10 is a sufficient training objective for localization.
    Eq. (2) and §2.2; standard in Dinomaly-style methods.
  • ad hoc to paper Selecting one annotated slice per exam and holding out slices ≥2 away from lesions yields an unbiased frame/pixel benchmark.
    Data preparation §3.1; risks optimistic or non-representative sampling of difficulty.
invented entities (1)
  • Panda pelvic Dinomaly pipeline (noisy MLP bottleneck + LA decoder + healthy-calibrated extent score s(x))
    purpose: Package frozen DINOv3 features into real-time unsupervised pelvic anomaly maps and slice flags without anomaly labels.
    Engineering assembly of existing UAD ideas specialized to uterine-cropped T2w MRI; not a new physical or statistical object.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Panda: Unsupervised Pelvic Anomaly Detection for Real-Time MR Imaging." pith.science (2026). https://pith.science/paper/CTEF5IW2

@misc{pith2026260724703,
  author       = {Pith},
  title        = {Pith review of: Panda: Unsupervised Pelvic Anomaly Detection for Real-Time MR Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CTEF5IW2}},
  note         = {Machine review of arXiv:2607.24703}
}
read the original abstract

Female pelvic diseases remain an under researched area characterized by often delayed diagnosis. While pelvic MRI offers superior soft-tissue contrast for diagnosis and image-guided procedures, real-time anomaly detection remains challenging due to physiological motion, tissue deformation, and instrument artifacts. Existing supervised approaches are impractical, as adverse events are rare, heterogeneous, and difficult to annotate. We present a Dinomaly-based unsupervised anomaly detection framework adapted for pelvic MRI that learns normative representations from healthy cases and flags deviations without requiring labels. Our approach leverages a frozen DINOv3 Vision Transformer encoder combined with a noisy MLP bottleneck and Linear Attention decoder to prevent identity mapping while maintaining computational efficiency. Anomalies are localized via per-token cosine distance between encoder and decoder representations, yielding spatial anomaly maps that provide immediate feedback at the scanner to support radiologist decision-making and adaptive protocol adjustment. Evaluated on a curated subset of the Uterine Myoma Dataset, the framework achieves a pixel-level AUROC of 88.06% and high specificity (95.45%) at frame level at 40.5 slices/s, meeting real-time clinical deployment requirements. The spatial anomaly maps and frame-level scores provide immediate, localized feedback at the scanner to support radiologist decision-making and adaptive protocol adjustment during active procedures.

Figures

Figures reproduced from arXiv: 2607.24703 by the authors.

Figure 1
Figure 1. (a) Schematic and T2w MRI pelvic illustration (pink—uterus; light [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Proposed framework for low-latency pelvic anomaly detection in MRI. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Representative results on Dtest. Each row presents from left to right: T2w input slice, GT, raw anomaly map, heatmap overlay on the input, and threshold￾based prediction maximizing the Dice across Dtest overlaid with GT on the input (yellow: predicted anomaly area, green: GT boundary, red: false positive) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on 2 representative cases: Input slice, GT mask, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 4 canonical work pages

  1. [1]

    Radiographics35(6), 1668–1676 (2015)

    Bruno, M.A., Walker, E.A., Abujudeh, H.H.: Understanding and confronting our mistakes: the epidemiology of error in radiology and strategies for error reduction. Radiographics35(6), 1668–1676 (2015)

  2. [2]

    In: CVPR

    Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., Liao, H.: Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. In: CVPR. pp. 20405–20415 (2025)

  3. [3]

    Medical physics49(4), 2774–2793 (2022)

    Hernando, D., Zhang, Y., Pirasteh, A.: Quantitative diffusion mri of the abdomen and pelvis. Medical physics49(4), 2774–2793 (2022)

  4. [4]

    Frontiers in Oncology 13, 1142947 (2023)

    Huang, P., Shang, J., Xu, Y., Hu, Z., Zhang, K., Dai, J., Yan, H.: Anomaly detec- tion in radiotherapy plans using deep autoencoder networks. Frontiers in Oncology 13, 1142947 (2023)

  5. [5]

    In: ICML

    Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: ICML. p. 5156. PMLR (2020)

  6. [6]

    Expert review of medical devices19(10), 763–778 (2022)

    Kilbride, B.F., Narsinh, K.H., Jordan, C.D., Mueller, K., Moore, T., Martin, A.J., Wilson,M.W.,Hetts,S.W.:Mri-guidedendovascularintervention:currentmethods and future potential. Expert review of medical devices19(10), 763–778 (2022)

  7. [7]

    IEEE signal processing magazine37(1), 128–140 (2020)

    Knoll, F., Hammernik, K., Zhang, C., Moeller, S., Pock, T., Sodickson, D.K., Ak- cakaya, M.: Deep-learning methods for parallel magnetic resonance imaging re- construction: A survey of the current approaches, trends, and issues. IEEE signal processing magazine37(1), 128–140 (2020)

  8. [8]

    Knupfer, A., Müller, J.P., Verdera, J.A., Fenske, M., Mathy, C.S., Tripathy, S., Arndt, S., May, M., Uder, M., Beckmann, M.W., Burghaus, S., Hutter, J.: Unsu- pervised anomaly detection of diseases in the female pelvis for real-time mr imaging (2026),https://arxiv.org/abs/2602.06179

Show all 28 references
  1. [9]

    Scientific reports10(1), 13710 (2020)

    Küstner, T., Fuin, N., Hammernik, K., Bustin, A., Qi, H., Hajhosseiny, R., Masci, P.G., Neji, R., Rueckert, D., Botnar, R.M., et al.: Cinenet: deep learning-based 3d cardiac cine mri reconstruction with multi-coil complex-valued 4d spatio-temporal convolutions. Scientific repo...

  2. [10]

    Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine237(1), 18–34 (2023)

    Liang,H.,Tse,Z.T.H.:Mrconditionalprostateinterventionsystemsandactuations review. Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine237(1), 18–34 (2023)

  3. [11]

    HuggingFace Hub, https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m(2025)

    Meta AI:facebook/dinov3-vits16-pretrain-lvd1689m. HuggingFace Hub, https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m(2025)

  4. [12]

    Knupfer et al

    Moynagh, M.R., Dowdy, S.C., Welch, B., Glaser, G.E., Schmitz, J.J., Jatoi, A., Langstraat, C.L., Block, M.S., Kurup, A.N., Kumar, A.: Image-guided tumor ab- lation in gynecologic oncology: review of interventional oncology techniques and 10 A. Knupfer et al. case examples high...

  5. [13]

    Müller, J.P.: Synthuterus roi (2026).https://doi.org/10.5281/ZENODO.18297879

  6. [14]

    Pan, H., Chen, M., Bai, W., Li, B., Zhao, X., Zhang, M., Zhang, D., Li, Y., Wang, H.: UMD (2023).https://doi.org/10.6084/m9.figshare.23541312.v3

  7. [15]

    Scientific Data11(1) (2024).https://doi.org/10.1038/s41597-024-03170-x

    Pan, H., Chen, M., Bai, W., Li, B., Zhao, X., Zhang, M., Zhang, D., Li, Y., Wang, H., Geng, H., Kong, W., Yin, C., Han, L., Lan, J., Zhao, T.: Large-scale uterine myoma mri dataset covering all figo types with pixel-level annotations. Scientific Data11(1) (2024).https://doi.or...

  8. [16]

    In: MICCAI

    Pinaya, W.H., Graham, M.S., Gray, R., Da Costa, P.F., Tudosiu, P.D., Wright, P., Mah, Y.H., MacKinnon, A.D., Teo, J.T., Jager, R., et al.: Fast unsupervised brain anomaly detection and segmentation with diffusion models. In: MICCAI. pp. 705–714 (2022)

  9. [17]

    The lancet oncology22(4), 538–547 (2021)

    Pötter, R., Tanderup, K., Schmid, M.P., Jürgenliemk-Schulz, I., Haie-Meder, C., Fokdal, L.U., Sturdza, A.E., Hoskin, P., Mahantshetty, U., Segedin, B., et al.: Mri- guided adaptive brachytherapy in locally advanced cervical cancer (embrace-i): a multicentre prospective cohort ...

  10. [18]

    Pattern Recognition p

    Rahmaniar, W., Suzuki, K.: Multi-ad: cross-domain unsupervised anomaly detec- tion for medical and industrial applications. Pattern Recognition p. 112486 (2025)

  11. [19]

    the innovative contribution of on-board hybrid mr imaging

    Romano, A., Votta, C., Nardini, M., Chiloiro, G., Panza, G., Boldrini, L., Cusumano, D., Galofaro, E., Placidi, L., Antonelli, M.V., et al.: Uterus motion analysis for radiotherapy planning optimization. the innovative contribution of on-board hybrid mr imaging. Clinical and T...

  12. [20]

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, ...

  13. [21]

    Medical image analysis63, 101693 (2020)

    Tajbakhsh, N., Jeyaseelan, L., Li, Q., Chiang, J.N., Wu, Z., Ding, X.: Embracing imperfect datasets: A review of deep learning solutions for medical image segmen- tation. Medical image analysis63, 101693 (2020)

  14. [22]

    Tripathy, S., Castro, N., May, M., Siegler, L., Story, L., Uder, M., Hutter, J.: Deep supervision attention u-net for segmentation of uterine zones: a multi-center study (2025).https://doi.org/10.58530/2025/0105, iSMRM 2025

  15. [23]

    Interventional Radiology9(3), 92–98 (2024)

    Uka, M., Matsui, Y., Iguchi, T., Matsushita, T., Tomita, K., Umakoshi, N., Kawa- bata, T., Yamada, M., Hiraki, T.: Magnetic resonance imaging guidance for per- cutaneous needle intervention. Interventional Radiology9(3), 92–98 (2024)

  16. [24]

    NeurIPS30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. NeurIPS30(2017)

  17. [25]

    Magnetic Resonance Imaging Clinics19(3), 547–566 (2011)

    Wasnik, A.P., Mazza, M.B., Liu, P.S.: Normal and variant pelvic anatomy on mri. Magnetic Resonance Imaging Clinics19(3), 547–566 (2011)

  18. [26]

    The British Journal of Radiology92(1098), 20180670 (2019)

    White, I.M., Scurr, E., Wetscherek, A., Brown, G., Sohaib, A., Nill, S., Oelfke, U., Dearnaley, D., Lalondrelle, S., Bhide, S.: Realizing the potential of magnetic reso- nance image guided radiotherapy in gynaecological and rectal cancer. The British Journal of Radiology92(109...

  19. [27]

    NeurIPS35, 4571 (2022)

    You, Z., Cui, L., Shen, Y., Yang, K., Lu, X., Zheng, Y., Le, X.: A unified model for multi-class anomaly detection. NeurIPS35, 4571 (2022)

  20. [28]

    Frontiers in Oncology15, 1597185 (2025)

    Yu, Y., Zhang, L., Sui, L., Zhang, H., Chen, L.: Application of mri-guided hys- teroscopic one-step resection in preserving the fertility of early endometrial cancer patients. Frontiers in Oncology15, 1597185 (2025)

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.