Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Interpretable Prediction of Lymph Node Metastasis in Rectal Cancer MRI Using Variational Autoencoders

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A variational autoencoder predicts lymph node metastasis in rectal cancer MRI at AUC 0.86, exceeding the pooled radiologist baseline.

desk verdict Plausible pilot study with a solid nested CV result, but the 'exceeds radiologists' claim rides on a test-set-selected model and an indirect comparison; worth reviewing after major revision. read the letter →

arxiv 2507.11638 v1 pith:42POBUWF submitted 2025-07-15 cs.CV cs.AIcs.LG

Benjamin Keel , Aaron Quyn , David Jayne , Maryam Mohsin , Samuel D. Relton This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords lymphnodemetastasisrectalcancervariationalautoencoderMRImultipleinstancelearningexplainableartificialintelligencestagingdeep
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

This paper argues that a variational autoencoder (VAE) can replace a large pre-trained CNN as the feature encoder for predicting lymph node metastasis in rectal cancer MRI, and that the resulting model is the first to exceed the pooled diagnostic accuracy of radiologists when lymph nodes are segmented by a non-specialist. The proposed model, VAE-MLP, encodes 2D lymph node patches into latent vectors and passes them through a two-stage MLP (multi-layer perceptron) classifier that combines individual node predictions with clinical features. On 168 untreated patients, it reports cross-validated area-under-the-curve (AUC) 0.86 ± 0.05, sensitivity 0.79 ± 0.06, and specificity 0.85 ± 0.05, compared with a meta-analytic radiologist baseline of 73% sensitivity and 74% specificity. The paper also claims the latent space is interpretable, clustering nodes by size and border irregularity and supporting smooth interpolation of node growth. If true, this would make accurate lymph node staging more accessible and less dependent on expert annotation.

What carries the argument

The central object is the VAE encoder, a generative model that compresses each lymph node patch into a low-dimensional latent vector by learning to reconstruct the patch; its weighted loss combines L1 reconstruction error, structural similarity (SSIM), and KL divergence with an annealing schedule. This latent representation replaces the features of a large pre-trained CNN and feeds into a multiple instance learning stack, where one MLP makes per-node predictions and a second MLP fuses those predictions with clinical features into the final patient diagnosis. Interpretability comes from Grad-CAM heatmaps on the encoder, k-means clustering of latent vectors that separates nodes by size and border irregularity, and latent-space interpolation along a node-growth direction.

What would settle it

Run the released pipeline on an independent multi-centre MRI cohort using the same preprocessing and non-specialist segmentations; if the AUC does not stay above the pooled radiologist baseline, the central claim is refuted. An audit of the Bayesian search logs would also settle it: if any test-set prediction influenced the choice of the final model, the headline numbers collapse.

Watch

Extended reading notes

Core claim

The central claim is that a generative feature encoder, rather than a classification CNN, is the right backbone for MRI-based lymph node staging in rectal cancer. A VAE was trained to reconstruct 32×32 lymph node patches, and its latent vectors were passed to a multiple instance learning classifier: a first MLP scores individual nodes, and a second MLP combines these scores with age, sex, T stage, and node size and shape features to produce a patient-level diagnosis. The best configuration reached 0.888 AUC with 92.3% sensitivity and 84.8% specificity on a 59-patient test set, and the five-fold nested cross-validation average was 0.858 AUC, 78.9% sensitivity, and 84.6% specificity. Against the pooled radiologist figures of 73% sensitivity and 74% specificity, the authors present this as the first demonstration that non-specialist segmentations can support accuracy above human readers. They also show that the latent space separates nodes by clinically relevant size and border irregularity and can simulate node growth by walking along the small-to-large direction in latent space.

Load-bearing premise

The whole comparison depends on the 35% test set being truly untouched by model selection; the hyperparameter searches were run to maximize test accuracy, so if any test information leaked into choosing the final model, the reported accuracy is too optimistic.

Editorial extensions

If this is right

  • Non-specialist lymph node segmentations, with only a fifth of scans reviewed by a consultant radiologist, may be sufficient for clinically usable staging, lowering the cost of building such models.
  • A VAE backbone can replace a large pre-trained CNN in this task without sacrificing accuracy, suggesting generative reconstruction is a viable pretraining strategy on small medical datasets.
  • The latent space separates nodes by size, shape, and border irregularity, giving clinicians a way to inspect what the model is using rather than treating it as a black box.
  • In the study cohort, the model could have flagged up to 92.3% of the 38 lymph-node-positive patients who were not referred for neo-adjuvant therapy, showing a concrete decision-support role.
  • Combining the maximum per-node probability with the patient-level MLP outperforms either aggregation alone, which is a simple but effective recipe for this multiple instance learning problem.

Reading between the lines

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

  • Beyond the paper: a direct external-validation study on another centre's MRI data is the natural next test; the current numbers come from one hospital, and scanner differences could shrink the reported gap over radiologists.
  • Beyond the paper: the 'exceeds radiologists' claim compares against a pooled meta-analytic baseline rather than a same-case reader study, so a head-to-head comparison with radiologists on the same scans would be a stronger test.
  • Beyond the paper: the latent-space growth interpolation hints at counterfactual explanations of the form 'this node would look like this if it became malignant,' which the paper lists as future work.
  • Beyond the paper: the same VAE-plus-MIL recipe could plausibly transfer to other cancers where nodal staging on MRI is similarly uncertain, but that extension is untested.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a VAE-based feature extractor combined with a two-stage MLP multiple-instance learning pipeline to predict lymph node metastasis in rectal cancer from T2-weighted MRI lymph-node patches and clinical variables. Using an in-house single-centre cohort of 168 patients with non-specialist lymph-node segmentations, the authors report a best held-out test AUC of 0.888, a five-fold nested cross-validated AUC of 0.858 ± 0.051, ablation comparisons against a CNN backbone and alternative fusion strategies, reconstruction metrics, Grad-CAM visualisations, latent-space clustering statistics, and an error analysis. The abstract and conclusions highlight sensitivity of 0.79–0.923 and claim that this is the first work to exceed radiologist diagnostic accuracy for this task using non-specialist segmentations.

Significance. If the evaluation were clean, this would be a useful contribution to a clinically important problem: the use of a lightweight VAE in place of a large pre-trained CNN is clearly motivated, the nested CV design is more rigorous than most prior single-split studies, and the public code and detailed ablation/error analyses support reproducibility. The interpretability analyses (Grad-CAM, latent-space clustering, and simulated LN growth) are valuable and give the work a distinctive angle. However, the central clinical claim — that the model exceeds radiologist accuracy — is currently supported only by test-set-selected point estimates compared against a meta-analysis on different patients, and the nested CV estimate itself needs a clear description of its selection procedure before it can be regarded as unbiased.

major comments (4)
  1. [§3.3 and Table 3] The 'best VAE-MLP' performance in Table 3 is not an unbiased estimate: §3.3 states that the Bayesian hyperparameter search aimed 'to improve the maximum test AUC', and the optimal hyperparameters in §4.2 are 'averaged over the top 10 performing models according to the maximum test AUC score'. Selecting the model with the highest test AUC on the same 59-patient test set that is then used to report the row inflates all point estimates (AUC 0.888, sensitivity 0.923, specificity 0.848). The abstract's cross-validated metrics are the only defensible headline numbers; the conclusion that the model 'exceeded' radiologist performance should not be based on the test-selected best model without a corrected evaluation protocol.
  2. [§3.3 (nested cross-validation)] The description of the nested CV is ambiguous and potentially contaminated: 'For each fold, the VAE with the highest test SSIM was selected from 5 candidates, followed by the MLP with the highest test AUC out of 10 candidates.' If 'test' refers to the held-out fold of the outer cross-validation, then model selection is using the evaluation data, and the reported CV AUC of 0.858 ± 0.051 is optimistic. If instead an inner validation split was used, this should be stated explicitly. Please clarify the split structure, and if necessary re-run the CV so that all selection decisions (including the VAE SSIM criterion) use only training/validation data.
  3. [§4.2 and Conclusion] The radiologist comparison is not statistically grounded. The pooled radiologist sensitivity of 73% and specificity of 74% come from the meta-analysis in [34], i.e., a different patient population, while no reader study was performed on the 59-patient test set. With only 13 pathologically positive patients, the 95% confidence interval for the best-model sensitivity of 0.923 is approximately 0.64–0.997, overlapping the meta-analytic estimate of 0.73; the cross-validated sensitivity (0.789 ± 0.064) is also not formally distinguishable. The claim 'first study to exceed the diagnostic accuracy of radiologists' requires a same-cohort reader study or at least a proper statistical comparison that accounts for the uncertainty in both estimates.
  4. [§4.3 and Table 3] The ablation text states that the CNN feature extractor 'had lower performance metrics than the VAE', but the CNN-MLP row in Table 3 reports a higher AUC (0.893 vs 0.888) and higher specificity (0.891 vs 0.848) than the best VAE-MLP model. If the intended comparison is to the cross-validated VAE-MLP (AUC 0.858), that should be stated explicitly. As written, the claim is contradicted by the table and should be corrected or qualified.
minor comments (5)
  1. [§3.3] The statement that 'the optimal hyperparameter ranges were discovered through three Bayesian searches with 200 model runs' is ambiguous relative to the earlier statement of a 'Bayesian search of the hyperparameter space with 200 runs'; please clarify whether each search had 200 runs or the total was 200.
  2. [Table 3 caption] The table caption should expand the model abbreviations (VAE-MLPDL, VAE-MLPP, VAE-MLPLN, VAE-MLPMax, VAE-MLPMLP) or add a note in the caption; the current table is hard to read without referring back to §4.3.
  3. [§3.5] The phrase 'this corresponds to a 94.8% sensitivity in detecting LNs' is not a standard use of the term sensitivity; please clarify what denominator is used (e.g., LNs found by the radiologist review) and consider using a different term.
  4. [§4.4] The 500 model states are selected using a per-epoch test AUC above 0.8; since this threshold uses the test set, the resulting error and uncertainty analysis may inherit selection bias. Please state this limitation or use validation-set selection for the ensemble.
  5. [§5] The statement that 'a high classification performance indicates that the annotations are accurate' is not a valid inference; classification performance against pathology does not establish segmentation quality, and this should be rephrased as a limitation.

Circularity Check

3 steps flagged · score 7.0 of 10

The headline 'exceeds radiologists' rests on the best model selected by test-AUC maximization, and the clustering evidence is self-definitional; the CV average provides partial but possibly contaminated independent support.

  1. fitted input called prediction [Section 3.3 (MLP Model Training) and Section 4.2 (Optimal Hyperparameters)]
    "These hyperparameters were optimised using a Bayesian search on WandB, aiming to improve the maximum test AUC. Classification model hyperparameters provided here are averaged over the top 10 performing models according to the maximum test AUC score."

    The reported best VAE-MLP row in Table 3 (AUC 0.888, sensitivity 0.923, specificity 0.848) is selected as the model with the maximum test AUC on the 35% held-out test set. The test-set AUC is therefore the selection objective, not an independent prediction. Reporting this same statistic as evidence that the model 'has exceeded the performance of the most robust studies' forces the headline result by construction; the hyperparameters were fitted to maximize the very metric later quoted as the model's performance.

  2. fitted input called prediction [Section 3.3 (MLP Model Training, nested cross-validation)]
    "For each fold, the VAE with the highest test SSIM was selected from 5 candidates, followed by the MLP with the highest test AUC out of 10 candidates."

    As written, model selection within each CV fold uses 'test SSIM' and 'test AUC' on the same held-out partition on which the fold is later evaluated; no inner validation split is described. Thus the reported CV mean (AUC 0.858±0.051) is partly a selection outcome rather than an unbiased estimate of generalization. If 'test' instead referred to an inner validation split this would not be circular, but the paper uses 'test' elsewhere to mean the held-out evaluation set and does not define an inner split.

1 more flagged steps
  1. self definitional [Section 3.4 (Post-Hoc Explainability) and Section 4.5 (Interpretability and Clustering)]
    "The number of clusters (k) was optimised to minimise the variability in lymph node size within the clusters. ... The standard deviation and ranges indicate that variability within the clusters is lower than the global variability, suggesting that patches are grouped based on similar LN size and shape."

    The clustering parameter k is explicitly fitted to minimize intra-cluster variability in LN size, so the later observation that clusters have lower intra-cluster size variability than the global population is true by construction. It cannot independently demonstrate that the VAE latent space 'groups LNs based on information relevant to the diagnosis including size' as claimed; the evidence sentence restates the optimization objective as a discovered property.

full rationale

The paper is not wholly circular: the abstract's headline metric (AUC 0.86 ± 0.05) is the five-fold nested CV average, which is less directly contaminated by the single test-set Bayesian search than the best-model row. However, the central 'first study to exceed radiologists' claim is argued in the conclusion from the best VAE-MLP row (0.89 AUC, 92.3% sensitivity, 84.8% specificity), and that row was produced by a Bayesian search explicitly 'aiming to improve the maximum test AUC', with optimal hyperparameters averaged over the top 10 models ranked by test AUC. Thus the headline performance reduces, by the paper's own description, to a selection statistic on the same test set. The nested CV description compounds this: per-fold selection uses 'highest test SSIM' and 'highest test AUC' without an inner validation split, so the CV mean may also be partly fitted. Separately, the interpretability claim that the latent space clusters by LN size is self-definitional because k was optimized to minimize intra-cluster size variability, and that same variability reduction is then presented as evidence. Self-citations [13,14] are present but not load-bearing for the derivation, and no uniqueness theorem or ansatz-smuggling pattern appears. Score 7 reflects a central claim that partially reduces to the selection criterion while some independent content (the architecture, and the CV average if the per-fold selection is actually performed on inner splits) remains.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper's performance claims rest on a large set of fitted hyperparameters (VAE loss weights, architecture sizes, MLP hyperparameters, decision threshold, synthetic patient count) optimized through Bayesian search. It also depends on domain assumptions about ground truth accuracy, segmentation quality, and dataset representativeness. There are no new physical entities.

free parameters (6)
  • VAE loss weights α, λ, γ, β = α=0.5, λ=4000, γ=3×batch size (per top-5 SSIM models); β and annealing not specified
    These weights balance reconstruction and KL divergence in Eq. 1 and were tuned via Bayesian search over 200 runs; they affect the learned latent representation.
  • VAE architecture: base and latent size scalar = base=20, latent size=20 (latent dim 400)
    Controls the number of feature maps and latent dimensionality; selected via Bayesian search.
  • Classification weight η = reported as 0.8 in Section 4.2, but Section 3.3 states η ∈ [0.5, 0.75]
    Weights the patient-level MLP prediction against the max individual LN probability; inconsistent ranges are a red flag.
  • Synthetic patient augmentation count and oversample ratio = 25 synthetic patients, oversample ratio 1.5
    Used to balance classes by augmenting patches and oversampling positives; hyperparameter optimized.
  • Prediction threshold = 0.436
    Threshold for binary classification; tuned to maximize test AUC.
  • MLP hyperparameters (learning rate, weight decay, batch size, dropout, layer sizes) = learning rate 6.59e-3, weight decay 0.16, batch size 128, dropout 0.4/0.3, hidden dims 2048/96
    Optimized via Bayesian search; affect classification performance.
assumptions (5)
  • domain assumption The pathology N stage (N0/N1/N2) is an accurate ground truth for lymph node metastasis.
    The paper uses post-operative pathological staging as ground truth, which is the clinical standard, but it is subject to sampling error during pathological examination.
  • domain assumption The lymph node segmentation masks, produced by a non-specialist and partially reviewed by a radiologist, are sufficiently accurate for the model.
    Section 3.5 describes the annotation process; only 20.2% of scans were reviewed by a consultant radiologist. Errors in segmentation could bias the extracted patches.
  • domain assumption The T stage clinical feature contributes independent information and is not a proxy that already encodes the label.
    T stage is used as a feature; the paper notes a simple T-stage rule achieves 0.605/0.723, so the model's performance may be driven by T stage.
  • domain assumption The single-center dataset is representative of the broader rectal cancer MRI population.
    All data are from one NHS trust over 2010-2022; no external validation cohort is used, so generalization is assumed.
  • ad hoc to paper Synthetic patients created by augmenting training patches are label-preserving and do not introduce distribution shift.
    Section 3.3: between 5 and 30 synthetic patients were created by applying augmentations to the patches and oversampling positive patients. This is a custom balancing method without validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Prediction of Lymph Node Metastasis in Rectal Cancer MRI Using Variational Autoencoders." pith.science (2026). https://pith.science/paper/42POBUWF

@misc{pith2026250711638,
  author       = {Pith},
  title        = {Pith review of: Interpretable Prediction of Lymph Node Metastasis in Rectal Cancer MRI Using Variational Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42POBUWF}},
  note         = {Machine review of arXiv:2507.11638}
}
read the original abstract

Effective treatment for rectal cancer relies on accurate lymph node metastasis (LNM) staging. However, radiological criteria based on lymph node (LN) size, shape and texture morphology have limited diagnostic accuracy. In this work, we investigate applying a Variational Autoencoder (VAE) as a feature encoder model to replace the large pre-trained Convolutional Neural Network (CNN) used in existing approaches. The motivation for using a VAE is that the generative model aims to reconstruct the images, so it directly encodes visual features and meaningful patterns across the data. This leads to a disentangled and structured latent space which can be more interpretable than a CNN. Models are deployed on an in-house MRI dataset with 168 patients who did not undergo neo-adjuvant treatment. The post-operative pathological N stage was used as the ground truth to evaluate model predictions. Our proposed model 'VAE-MLP' achieved state-of-the-art performance on the MRI dataset, with cross-validated metrics of AUC 0.86 +/- 0.05, Sensitivity 0.79 +/- 0.06, and Specificity 0.85 +/- 0.05. Code is available at: https://github.com/benkeel/Lymph_Node_Classification_MIUA.

Figures

Figures reproduced from arXiv: 2507.11638 by the authors.

Figure 1
Figure 1. VAE-MLP model architecture The encoder has 6 blocks of 2D convolutional layers with 2D batch normali￾sation [11], and Gaussian Error Linear Unit (GELU) activation [8]. The output of the encoder is used in two separate convolutional layers for the latent vec￾tors, mean (µ) and log variance log(σ 2 )  , forming a latent space of LN feature representations. The decoder takes a sample from the n-dimensional Gaussian di… view at source ↗
Figure 1
Figure 1. The MLP layers included 1D Batch Normalisation, dropout and GELU [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Observe that the key structures are captured very well although some of [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figures from the paper (3 more)
Figure 2
Figure 2. Figure 2: MRI lymph node patches and VAE reconstructions [PITH_FULL_IMAGE:figures/full_fig_p009_2.png]
Figure 3
Figure 3. Figure 3: Grad-CAM heatmaps showing gradients from the last spatial layer of the [PITH_FULL_IMAGE:figures/full_fig_p012_3.png]
Figure 4
Figure 4. Figure 4: Lymph node growth direction, with the original image in the centre and [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 24 canonical work pages

  1. [34]

    Frontiers in Oncology11(July) (2021)

    Zhuang, Z., Zhang, Y., Wei, M., Yang, X., Wang, Z.: Magnetic Resonance Imaging Evaluation of the Accuracy of Various Lymph Node Staging Criteria in Rectal Cancer: A Systematic Review and Meta-Analysis. Frontiers in Oncology11(July) (2021). https://doi.org/10.3389/fonc.2021.709070

    work page arXiv 2021

  2. [1]

    Nhs digital cancer survival in england, cancers diagnosed 2016 to 2020 (2023), https://digital.nhs.uk/data-and-information/publications/statistical/ cancer-survival-in-england/cancers-diagnosed-2016-to-2020-followed-up-to-2021/

    work page 2023

  3. [2]

    wandb.com/

    Biewald, L.: Experiment tracking with weights and biases (2020), https://www. wandb.com/

    work page 2020

  4. [3]

    Journal of Clinical Medicine11(9), 1–26 (2022)

    Borgheresi, A., De Muzio, F., Agostini, A., Ottaviani, L., Bruno, A., et al.: Lymph Nodes Evaluation in Rectal Cancer: Where Do We Stand and Future Perspec- tive. Journal of Clinical Medicine11(9), 1–26 (2022). https://doi.org/10.3390/ jcm11092599

    work page 2022

  5. [4]

    Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., et al.: Monai: An open- source framework for deep learning in healthcare (2022), https://arxiv.org/abs/ 2211.02701 Interpretable LNM prediction 15

    arXiv 2022

  6. [5]

    In: 2009 IEEE Conference on Computer Vi- sion and Pattern Recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE Conference on Computer Vi- sion and Pattern Recognition. pp. 248–255 (2009). https://doi.org/10.1109/CVPR. 2009.5206848

    arXiv 2009

  7. [6]

    Variational Classification

    Dhuliawala, S., Sachan, M., Allen, C.: Variational Classification pp. 1–17 (2023), http://arxiv.org/abs/2305.10406

    work page Pith review arXiv 2023

  8. [7]

    Magnetic Resonance Imaging30(9), 1323–1341 (November 2012), https://www

    Fedorov, A., Beichel, R., Kalpathy-Cramer, J., Finet, J., Fillion-Robin, J.C., et al.: 3d slicer as an image computing platform for the quantitative imaging network. Magnetic Resonance Imaging30(9), 1323–1341 (November 2012), https://www. ncbi.nlm.nih.gov/pmc/articles/PMC3466397/pdf/nihms383480.pdf

    work page 2012

Show all 34 references
  1. [8]

    https://doi

    Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus) (2016). https://doi. org/10.48550/ARXIV.1606.08415, https://arxiv.org/abs/1606.08415

    2016 doi arXiv

  2. [9]

    RadioGraphics39(2), 367–387 (2019)

    Horvat, N., Carlos Tavares Rocha, C., Clemente Oliveira, B., Petkovska, I., Gollub, M.J.: Mri of rectal cancer: Tumor staging, imaging techniques, and management. RadioGraphics39(2), 367–387 (2019). https://doi.org/10.1148/rg.2019180114, https://doi.org/10.1148/rg.2019180114, ...

    2019 doi

  3. [10]

    Bioinformatics Advances3(1),vbac100(012023).https://doi.org/10.1093/bioadv/vbac100,https: //doi.org/10.1093/bioadv/vbac100

    Hsu, T.C., Lin, C.: Learning from small medical data—robust semi-supervised cancer prognosis classifier with bayesian variational autoencoder. Bioinformatics Advances3(1),vbac100(012023).https://doi.org/10.1093/bioadv/vbac100,https: //doi.org/10.1093/bioadv/vbac100

    doi

  4. [11]

    CoRR (2015), http://arxiv.org/abs/1502.03167

    Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. CoRR (2015), http://arxiv.org/abs/1502.03167

    2015 arXiv

  5. [12]

    Journal of Artificial Intelligence and Technology3(4), 195–204 (2023)

    Kaur, G., Keshta, I., Shabaz, M., Batra, H.S., Vijaya Sagar, T., et al.: Rectal Cancer Prediction and Performance Based on Intelligent Variational Autoencoders Machine Using Deep Learning on CDAS Dataset. Journal of Artificial Intelligence and Technology3(4), 195–204 (2023). h...

    2023

  6. [13]

    In: British Machine Vision Conference (BMVC) (2023)

    Keel, B., Quyn, A., Jayne, D., Relton, S.D.: Variational autoencoders for feature exploration and malignancy prediction of lung lesions. In: British Machine Vision Conference (BMVC) (2023). https://doi.org/10.48550/arXiv.2311.15719, https:// doi.org/10.48550/arXiv.2311.15719

    2023 doi arXiv

  7. [14]

    BMJ Open14(12) (2024)

    Keel, B., Quyn, A., Jayne, D., Relton, S.D.: State-of-the-art performance of deep learning methods for pre-operative radiologic staging of colorectal cancer lymph node metastasis: a scoping review. BMJ Open14(12) (2024). https://doi.org/10. 1136/bmjopen-2024-086896, https://bm...

    2024

  8. [15]

    arXiv (2013)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv (2013). https: //doi.org/10.48550/ARXIV.1312.6114, https://arxiv.org/abs/1312.6114

    2013 doi arXiv

  9. [16]

    The Annals of Math- ematical Statistics22(1), 79–86 (1951), http://www.jstor.org/stable/2236703

    Kullback, S., Leibler, R.A.: On information and sufficiency. The Annals of Math- ematical Statistics22(1), 79–86 (1951), http://www.jstor.org/stable/2236703

    1951

  10. [17]

    Quantitative Imaging in Medicine and Surgery11(6), 2477–2485 (2021)

    Li, J., Zhou, Y., Wang, P., Zhao, H., Wang, X., et al.: Deep transfer learning based on magnetic resonance imaging can improve the diagnosis of lymph node metastasis in patients with rectal cancer. Quantitative Imaging in Medicine and Surgery11(6), 2477–2485 (2021). https://do...

    2021 doi

  11. [18]

    International Jour- nal of Radiation Research21(2) (2023)

    Liu, J., Sun, L., Lu, X., Geng, Y., Zhang, Z.: Development and validation of a stacking nomogram for predicting regional lymph node metastasis status in rectal cancer via deep learning and hand-crafted radiomics. International Jour- nal of Radiation Research21(2) (2023). https...

    2023 doi

  12. [19]

    Keel et al

    MacQueen, J.: Some methods for classification and analysis of multivariate obser- vations (1967) 16 B. Keel et al

    1967

  13. [20]

    Diseases of the Colon and Rectum66(12), E1246–E1253 (2023)

    Ozaki, K., Kurose, Y., Kawai, K., Kobayashi, H., Itabashi, M., et al.: Development of a Diagnostic Artificial Intelligence Tool for Lateral Lymph Node Metastasis in Advanced Rectal Cancer. Diseases of the Colon and Rectum66(12), E1246–E1253 (2023). https://doi.org/10.1097/DCR....

    2023 doi

  14. [21]

    Computer Methods and Programs in Biomedicine p

    Pérez-García, F., Sparks, R., Ourselin, S.: TorchIO: a Python library for effi- cient loading, preprocessing, augmentation and patch-based sampling of medi- cal images in deep learning. Computer Methods and Programs in Biomedicine p. 106236 (2021). https://doi.org/https://doi....

    2021

  15. [22]

    Rais, K., Amroune, M., Benmachiche, A., Haouam, M.Y.: Exploring variational autoencoders for medical image generation: A comprehensive study (2024), https: //arxiv.org/abs/2411.07348

    2024 arXiv

  16. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    Redekop, E., Pleasure, M., Wang, Z., Sarma, K.V., Kinnaird, A., et al.: Code- book vq-vae approach for prostate cancer diagnosis using multiparametric mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 2365–2372 (June 2024)

    2024

  17. [24]

    International Journal of Computer Vision128(2), 336–359 (oct 2019)

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- CAM: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision128(2), 336–359 (oct 2019). https:// doi.org/10.1007/s11263-019-01228-7

    2019 doi

  18. [25]

    https://www.slicer.org/ (2024), accessed: 2024-10-24

    Slicer, D.: 3d slicer: An open source software platform for medical image computing and research. https://www.slicer.org/ (2024), accessed: 2024-10-24

    2024

  19. [26]

    IEEE Access12, 195658– 195670 (2024)

    Sreelekshmi, V., Pavithran, K., Nair, J.J.: Unleashing the power of hierarchical variational autoencoder for predicting breast cancer. IEEE Access12, 195658– 195670 (2024). https://doi.org/10.1109/ACCESS.2024.3518612

    2024

  20. [27]

    Journal of Clinical Oncology29(23), 3163–3172 (2011)

    Valentini, V., van Stiphout, R.G., Lammering, G., Gambacorta, M.A., Barba, M.C., et al.: Nomograms for predicting local recurrence, distant metastases, and overall survival for patients with locally advanced rectal cancer on the ba- sis of european randomized clinical trials. ...

    2011 doi

  21. [28]

    European Radiology pp

    Wan, L., Hu, J., Chen, S., Zhao, R., Peng, W., et al.: Prediction of lymph node metastasis in stage T1–2 rectal cancers with MRI-based deep learning. European Radiology pp. 3638–3646 (2023). https://doi.org/10.1007/s00330-023-09450-1, https://doi.org/10.1007/s00330-023-09450-1

    2023 doi

  22. [29]

    IEEE Access12, 90535–90551 (2024)

    Wang,J.,Li,J.,Wang,R.,Zhou,X.:Vae-drivenmultimodalfusionforearlycardiac disease detection. IEEE Access12, 90535–90551 (2024). https://doi.org/10.1109/ ACCESS.2024.3420444

    2024

  23. [30]

    IEEE Transactions on Image Processing 13(4), 600–612 (2004), https://doi.org/10.1109/TIP.2003.819861

    Wang, Z., Bovik, A.C., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004), https://doi.org/10.1109/TIP.2003.819861

    2004

  24. [31]

    Radiology: Artificial Intelligence6(2) (2024)

    Xia, W., Li, D., He, W., Pickhardt, P.J., Jian, J., et al.: Multicenter Evaluation of a Weakly Supervised Deep Learning Model for Lymph Node Diagnosis in Rectal Cancer at MRI. Radiology: Artificial Intelligence6(2) (2024). https://doi.org/10. 1148/ryai.230152

    2024

  25. [32]

    2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) pp

    Xie, M., Zhang, Y., Li, X., Mao, Y., Zou, X., Zhang, H.: Predicting Lymph Node Metastasis of Colorectal Cancer in CT Scans Using Attention-Based Multi- ple Instance Learning. 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) pp. 2695–2701 (2023). http...

    2023

  26. [33]

    In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreason- able effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586–595 (2018). https://doi.org/10.1109/CVPR.2018.00068

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.