pith. machine review for the scientific record. sign in

arxiv: 2605.05522 · v1 · submitted 2026-05-06 · 📡 eess.IV · cs.CV

Recognition: unknown

Tumor-aware augmentation with task-guided attention analysis improves rectal cancer segmentation from magnetic resonance images

Aditi Iyer, Aditya Apte, Aneesh Rangnekar, Eve LoCastro, Harini Veeraraghavan, Iva Petkovska, Jesse Joshua Smith, Joao Miranda, Joseph Deasy, Julio Garcia-Aguilar, Marc J Gollub, Natally Horvat, Paul Romesser, Revathi Ravella, Samir Alrayess, Stephanie Chahwan

Authors on Pith no claims yet

Pith reviewed 2026-05-08 15:17 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords rectal cancer segmentationCT-to-MRI transfertransformer pretrainingattention dilutionmedical image augmentationhierarchical transformers
0
0 comments X

The pith

Tumor-aware augmentation and anisotropic cropping restore token efficiency in CT-pretrained transformers for rectal MRI segmentation.

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

The paper demonstrates that common assumptions about dataset adaptation and cross-modality transfer break down when applying CT-pretrained hierarchical transformers to rectal cancer segmentation in MRI scans. Two specific failure modes appear: zero-padding creates inefficient token usage by diverting attention to uninformative regions, and pretrained features show limited adaptation despite fine-tuning. Using an attention dilution index to quantify padding effects and centered kernel alignment to track feature reuse, the authors test two interventions on SMIT and Swin UNETR backbones. Tumor-aware augmentation increases coverage of tumor appearance variations while anisotropic cropping eliminates wasteful padding. These changes raise detection rates to 224 out of 247 cases for one model and 219 out of 247 for the other on the same rectal MRI test set.

Core claim

Mechanistic analysis of attention dilution and feature reuse shows that zero-padding and ineffective adaptation cause accuracy loss in CT-to-MRI transfer; tumor-aware augmentation plus anisotropic cropping directly mitigate both issues and improve detection rates to 90.7 percent and 88.7 percent on identical rectal MRI datasets for SMIT and Swin UNETR respectively.

What carries the argument

Attention dilution index (ADI), an entropy-based metric that measures how much attention is diverted to zero-padded tokens, used together with centered kernel alignment (CKA) to assess feature reuse across modalities.

Load-bearing premise

The failure modes of token inefficiency and poor feature adaptation are the main reasons for degraded transfer performance and can be fixed by the proposed augmentation and cropping steps without creating selection bias.

What would settle it

Applying the same tumor-aware augmentation and anisotropic cropping to a model trained from scratch on MRI data alone produces no detection-rate improvement over standard fine-tuning.

Figures

Figures reproduced from arXiv: 2605.05522 by Aditi Iyer, Aditya Apte, Aneesh Rangnekar, Eve LoCastro, Harini Veeraraghavan, Iva Petkovska, Jesse Joshua Smith, Joao Miranda, Joseph Deasy, Julio Garcia-Aguilar, Marc J Gollub, Natally Horvat, Paul Romesser, Revathi Ravella, Samir Alrayess, Stephanie Chahwan.

Figure 1
Figure 1. Figure 1: Tumor-aware intensity augmentation applied to representative training cases. Each row shows an axial slice with manual delineation (green), followed by three augmented variants in which tumor-region intensities are randomly scaled and shifted. perturbations that broadened tumor appearance without in￾troducing implausible contrast inversions or overwhelming the underlying anatomy. The perturbation was appli… view at source ↗
Figure 2
Figure 2. Figure 2: Rows A–C show representative high-contrast tumors (Cluster-B), while rows D–F show low-contrast tumors (Cluster-D). Each row presents the MRI slice with manual delineation (green), followed by predictions from SMIT and Swin UNETR across the three configurations. Surface DSC (sDSC) is shown in the lower-right corner of each prediction panel. Missing values indicate that the tumor was not detected. The final… view at source ↗
Figure 3
Figure 3. Figure 3: Representative challenging tumors under the ACT configuration. Each row shows an axial MRI slice with reference contour (green), followed by predictions from SMIT-ACT and Swin UNETR-ACT (orange). Surface DSC (sDSC) is shown in the prediction panels; ‘Not detected’ indicates failed tumor detection. to the pretrained representations across most layers, indi￾cating that tumor-aware augmentation primarily infl… view at source ↗
Figure 4
Figure 4. Figure 4: Attention dilution index (ADI) analysis linking hierarchical depth, padding fraction, and segmentation accuracy. (A–B) Mean ADI per stage for SMIT and Swin UNETR; the shaded region highlights the S0–S2 window artifact in Swin UNETR. (C) Mean Stage-3 ADI illustrating the effect of anisotropic cropping across both backbones. (D) Per-scan Stage-3 ADI as a function of padding fraction (𝑝𝑓), with per-backbone S… view at source ↗
Figure 5
Figure 5. Figure 5: Two tumor subtypes identified by unsupervised clus￾tering were consistent across cohorts. Tumor-to-background contrast versus boundary gradient for the training (A) and held-out test set (B). Cluster-B (blue) tumors show higher contrast and sharper boundaries than Cluster-D (orange). Donut insets show subtype proportions. Note that one contrast outlier was excluded from panel B. Through mechanistic analysi… view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise feature similarity to the pretrained backbone measured using linear CKA. Diagonal CKA across transformer blocks for (a) SMIT and (b) Swin UNETR under different configurations. 0.00 0.25 0.50 0.75 1.00 1.25 1.50 Training loss a 0.50 0.55 0.60 0.65 Validation Dice b Epoch 0.25 0.50 0.75 1.00 1.25 1.50 Training loss c Epoch 0.500 0.525 0.550 0.575 0.600 0.625 0.650 Validation Dice d Base Tumor-awar… view at source ↗
Figure 8
Figure 8. Figure 8: Layer-wise feature similarity between randomly initialized (Scratch) and CT-pretrained (CT-Init.) models mea￾sured using diagonal linear CKA across transformer blocks for SMIT and Swin UNETR. evaluation was performed using retrospectively collected data from a single institution and primarily GE Health￾care scanners, limiting assessment of robustness across vendors, field strengths, acquisition protocols, … view at source ↗
read the original abstract

Pretraining on large-scale datasets has been shown to improve transformer generalizability, even for out-of-domain (OOD) modalities and tasks. However, two common assumptions often fail under OOD transfer: that downstream datasets can be adapted to the fixed input geometry of pretrained models and that pretrained representations transfer effectively across imaging modalities. We show that these assumptions break down through two interacting failure modes in CT-to-MRI transfer: inefficient token usage caused by zero-padding to match pretrained input dimensions and ineffective feature adaptation. These failures led to accuracy degradation despite extensive fine-tuning. We investigated these failure modes using two CT-pretrained hierarchical shifted-window transformer backbones, SMIT and Swin UNETR, pretrained with different objectives and datasets. Mechanistic analysis introduced an attention dilution index (ADI), an entropy-based metric quantifying attention diverted toward uninformative padding tokens, and centered kernel alignment (CKA) to measure feature reuse in MRI tasks. ADI increased with zero-padding, while high feature reuse did not necessarily correspond to improved accuracy. To mitigate these issues, we introduced two interventions: a tumor-aware augmentation strategy to improve tumor appearance heterogeneity coverage and an anisotropic cropping strategy to restore token efficiency. Fine-tuning on identical rectal MRI datasets improved detection rates to 224/247 (90.7%) for SMIT and 219/247 (88.7%) for Swin UNETR, demonstrating improved robustness under CT-to-MRI transfer. This study is among the first to examine when pretrained transformers fail to transfer effectively across imaging modalities and how simple mitigation strategies, motivated by mechanistic analysis of datasets, can reduce transfer limitations while improving robustness and MRI detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper examines failure modes in transferring CT-pretrained hierarchical shifted-window transformers (SMIT and Swin UNETR) to rectal MRI segmentation tasks. It identifies two interacting issues—inefficient token usage from zero-padding to match pretrained input sizes and ineffective feature adaptation across modalities—using a new Attention Dilution Index (ADI) based on entropy and centered kernel alignment (CKA) for mechanistic analysis. The authors propose tumor-aware augmentation to increase tumor heterogeneity coverage and anisotropic cropping to restore token efficiency, reporting post-intervention detection rates of 224/247 (90.7%) for SMIT and 219/247 (88.7%) for Swin UNETR after fine-tuning on rectal MRI data, claiming improved robustness under CT-to-MRI transfer.

Significance. If the central improvements can be causally attributed to the proposed interventions via proper controls, the work would offer useful insights into cross-modality transfer limitations for medical vision transformers and practical, low-cost mitigation strategies. This could have moderate significance for improving segmentation robustness in rectal cancer MRI, where pretrained models are increasingly used but often degrade on OOD data.

major comments (3)
  1. [Abstract and Results] Abstract and Results: The final detection rates of 224/247 (90.7%) for SMIT and 219/247 (88.7%) for Swin UNETR are presented as evidence of improved robustness, but the corresponding rates from the prior 'extensive fine-tuning' (which the text states exhibited degradation) are not reported, preventing quantification of the actual improvement magnitude attributable to the interventions.
  2. [Methods/Results] Experimental design (implied in Methods/Results): No ablation studies are provided that apply tumor-aware augmentation alone, anisotropic cropping alone, or neither intervention (beyond the baseline extensive fine-tuning), which is load-bearing because the central claim attributes the accuracy gains specifically to these two strategies mitigating token inefficiency and feature adaptation failures.
  3. [Analysis] Analysis section: ADI and CKA are introduced and described as increasing with zero-padding and measuring feature reuse, respectively, but the manuscript does not include quantitative correlation (e.g., regression or per-case analysis) between ADI/CKA values and the final accuracy deltas, weakening the mechanistic justification for the mitigations.
minor comments (2)
  1. [Abstract] The abstract uses 'detection rates' for what is described as a segmentation task; clarify the exact metric (e.g., whether it is tumor presence detection within segmented volumes or a proxy for segmentation performance) to avoid ambiguity.
  2. [Methods] The definition and computation of the Attention Dilution Index (ADI) should be formalized with an equation in the main text rather than described only qualitatively, to allow reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which have helped us identify areas where the manuscript can be strengthened. We address each major comment below and have revised the manuscript accordingly to provide clearer quantification of improvements, additional controls, and stronger mechanistic evidence.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results: The final detection rates of 224/247 (90.7%) for SMIT and 219/247 (88.7%) for Swin UNETR are presented as evidence of improved robustness, but the corresponding rates from the prior 'extensive fine-tuning' (which the text states exhibited degradation) are not reported, preventing quantification of the actual improvement magnitude attributable to the interventions.

    Authors: We agree that explicit baseline rates are needed for direct comparison. The manuscript text notes degradation under extensive fine-tuning alone, and the underlying per-model detection rates from that condition are available from our experiments. In the revised version, we will report these baseline rates alongside the post-intervention figures in both the abstract and results sections to quantify the improvement magnitude. revision: yes

  2. Referee: [Methods/Results] Experimental design (implied in Methods/Results): No ablation studies are provided that apply tumor-aware augmentation alone, anisotropic cropping alone, or neither intervention (beyond the baseline extensive fine-tuning), which is load-bearing because the central claim attributes the accuracy gains specifically to these two strategies mitigating token inefficiency and feature adaptation failures.

    Authors: We acknowledge that separate ablations would strengthen attribution of gains to each intervention. Our original design emphasized the combined application motivated by the interacting failure modes identified via ADI and CKA. In the revision, we will add ablation experiments applying tumor-aware augmentation alone and anisotropic cropping alone, reporting their individual effects on detection rates, ADI, and CKA to isolate contributions. revision: yes

  3. Referee: [Analysis] Analysis section: ADI and CKA are introduced and described as increasing with zero-padding and measuring feature reuse, respectively, but the manuscript does not include quantitative correlation (e.g., regression or per-case analysis) between ADI/CKA values and the final accuracy deltas, weakening the mechanistic justification for the mitigations.

    Authors: We appreciate this point on strengthening the mechanistic link. The current analysis shows ADI rising with padding and CKA patterns for feature reuse, but lacks explicit correlation to accuracy. In the revised manuscript, we will add per-case scatter plots, Pearson/Spearman correlations, and regression analysis between ADI/CKA values and segmentation accuracy deltas across cases to provide quantitative support for the proposed mitigations. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical study with independent metrics and reported rates

full rationale

The paper is an empirical transfer-learning study. It defines ADI (entropy-based attention metric) and CKA externally, applies tumor-aware augmentation and anisotropic cropping as interventions, and reports concrete detection rates (224/247, 219/247) on rectal MRI data. No equations, fitted parameters, or predictions are shown to reduce by construction to the inputs; the central claim rests on experimental outcomes and mechanistic analysis rather than self-referential definitions or self-citation chains. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that the proposed interventions address the diagnosed transfer failures and that the reported detection rates reflect genuine robustness gains rather than dataset-specific effects.

axioms (1)
  • domain assumption Pretrained hierarchical shifted-window transformers on CT data can be meaningfully fine-tuned for MRI segmentation tasks
    Invoked throughout the transfer learning setup described in the abstract.
invented entities (1)
  • Attention Dilution Index (ADI) no independent evidence
    purpose: Entropy-based metric to quantify attention diverted to uninformative padding tokens
    Newly introduced diagnostic tool for analyzing transformer behavior under zero-padding.

pith-pipeline@v0.9.0 · 5673 in / 1327 out tokens · 41153 ms · 2026-05-08T15:17:18.738165+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    Prospective validation of a low rectal cancer magnetic resonance imaging staging system and development of a local recurrence risk stratification model: the mercury ii study

    Battersby,N.J.,How,P.,Moran,B.,Stelzner,S.,West,N.P.,Branagan, G., Strassburg, J., Quirke, P., Tekkis, P., Pedersen, B.G., et al., 2016. Prospective validation of a low rectal cancer magnetic resonance imaging staging system and development of a local recurrence risk stratification model: the mercury ii study. Annals of surgery 263, 751–760

    2016

  2. [2]

    Beets-Tan, R.G., Lambregts, D.M., Maas, M., Bipat, S., Barbaro, B., Curvo-Semedo, L., Fenlon, H.M., Gollub, M.J., Gourtsoyianni, S., Halligan, S., et al., 2018. Magnetic resonance imaging for clinical management of rectal cancer: updated recommendations from the 2016 european society of gastrointestinal and abdominal radiology (esgar) consensus meeting. E...

    2018

  3. [3]

    MONAI: An open-source framework for deep learning in healthcare

    Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey,B.,Myronenko,A.,Zhao,C.,Yang,D.,etal.,2022. MONAI: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701

    work page internal anchor Pith review arXiv 2022

  4. [4]

    Charbel, C., Kwok, H.C., Miranda, J., Zheng, J., El Homsi, M., El Amine, M.A., Chhabra, S., Danilova, S., Gangai, N., Petkovska, I., Capanu, M., Vanguri, R.S., Chakraborty, J., Horvat, N., 2025. Reliability of rectal mri radiomic features: Comparing rectal mri radiomic features across reader expertise levels, image segmenta- tion technique, and timing of ...

    work page doi:10.1016/j.ejrad.2025.112019 2025

  5. [5]

    Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers

    Chen, J., Mei, J., Li, X., Lu, Y., Yu, Q., Wei, Q., Luo, X., Xie, Y., Adeli, E., Wang, Y., et al., 2024. Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers. Medical Image Analysis

    2024

  6. [6]

    Med3d: Transfer learning for 3d medical image analysis

    Chen, S., Ma, K., Zheng, Y., 2019. Med3d: Transfer learning for 3d medical image analysis. arXiv preprint arXiv:1904.00625

    work page arXiv 2019

  7. [7]

    Algorithms for learningkernelsbasedoncenteredalignment.TheJournalofMachine Learning Research

    Cortes, C., Mohri, M., Rostamizadeh, A., 2012. Algorithms for learningkernelsbasedoncenteredalignment.TheJournalofMachine Learning Research

    2012

  8. [8]

    Vision transformersneedregisters,in:InternationalConferenceonLearning Representations

    Darcet, T., Oquab, M., Mairal, J., Bojanowski, P., 2024. Vision transformersneedregisters,in:InternationalConferenceonLearning Representations

    2024

  9. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Representations (ICLR)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N., 2021. An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Representations (ICLR)

    2021

  10. [10]

    Artificial intelligence–based technology for semi-automated segmentation of rectal cancer using high-resolution mri

    Hamabe,A.,Ishii,M.,Kamoda,R.,Sasuga,S.,Okuya,K.,Okita,K., Akizuki, E., Sato, Y., Miura, R., Onodera, K., et al., 2022. Artificial intelligence–based technology for semi-automated segmentation of rectal cancer using high-resolution mri. PLoS One 17, e0269931

    2022

  11. [11]

    Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.,

  12. [12]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images, in: International MICCAI brainlesion workshop, Springer. pp. 272–284

  13. [13]

    Horvat, N., Veeraraghavan, H., Khan, M., Blazic, I., Zheng, J., Ca- panu, M., Sala, E., Garcia-Aguilar, J., Gollub, M.J., Petkovska, I.,

  14. [14]

    Radiology 287, 833– 843

    Mr imaging of rectal cancer: radiomics analysis to assess treatment response after neoadjuvant therapy. Radiology 287, 833– 843

  15. [15]

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.,

  16. [16]

    Nature Methods

    nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods

  17. [17]

    Self-supervised learning improves robustness of deep learning lung tumor segmen- tation models to ct imaging differences

    Jiang, J., Rangnekar, A., Veeraraghavan, H., 2025. Self-supervised learning improves robustness of deep learning lung tumor segmen- tation models to ct imaging differences. Medical Physics 52, 1573– 1588

    2025

  18. [18]

    Jiang,J.,Tyagi,N.,Tringale,K.,Crane,C.,Veeraraghavan,H.,2022. Self-supervised3danatomysegmentationusingself-distilledmasked image transformer (SMIT), in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer

    2022

  19. [19]

    Semi-automatic tumor segmentation of rectal cancer based on functional magnetic resonance imaging

    Knuth, F., Groendahl, A.R., Winter, R.M., Torheim, T., Negård, A., Holmedal,S.H.,Bakke,K.M.,Meltzer,S.,Futsæther,C.M.,Redalen, K.R., 2022. Semi-automatic tumor segmentation of rectal cancer based on functional magnetic resonance imaging. Physics and imag- ing in radiation oncology 22, 77–84

    2022

  20. [20]

    Similarity of neural network representations revisited, in: International conference on machine learning, PMlR

    Kornblith, S., Norouzi, M., Lee, H., Hinton, G., 2019. Similarity of neural network representations revisited, in: International conference on machine learning, PMlR. pp. 3519–3529

    2019

  21. [21]

    Fully automated segmentation and radiomics feature extraction of hypopharyngeal cancer on mri using deep learning

    Lin, Y.C., Lin, G., Pandey, S., Yeh, C.H., Wang, J.J., Lin, C.Y., Ho, T.Y., Ko, S.F., Ng, S.H., 2023. Fully automated segmentation and radiomics feature extraction of hypopharyngeal cancer on mri using deep learning. European Radiology 33, 6548–6556

    2023

  22. [22]

    Is it time to replace cnns with transformers for medical images? arXiv preprint arXiv:2108.09038

    Matsoukas, C., Haslum, J.F., Söderberg, M., Smith, K., 2021. Is it time to replace cnns with transformers for medical images? arXiv preprint arXiv:2108.09038 . A. Rangnekar et al.:Preprint submitted to ElsevierPage 12 of 13 CT-to-MRI pretraining transfer for rectal tumor segmentation

    work page arXiv 2021

  23. [23]

    Matsoukas, C., Haslum, J.F., Sorkhei, M., Söderberg, M., Smith, K.,

  24. [24]

    9225–9234

    Whatmakestransferlearningworkformedicalimages:Feature reuse&otherfactors,in:ProceedingsoftheIEEE/CVFconferenceon computer vision and pattern recognition, pp. 9225–9234

  25. [25]

    V-net: Fully convolu- tional neural networks for volumetric medical image segmentation, in: 2016 fourth international conference on 3D vision (3DV), Ieee

    Milletari, F., Navab, N., Ahmadi, S.A., 2016. V-net: Fully convolu- tional neural networks for volumetric medical image segmentation, in: 2016 fourth international conference on 3D vision (3DV), Ieee. pp. 565–571

    2016

  26. [26]

    Machado, F.A., Chakraborty, J., Pandini, R.V., Saraiva, S., Nahas, C.S.R., Nahas, S.C.,Nomura,C.H.,2023

    Miranda, J., Horvat, N., Assuncao Jr, A.N., de M. Machado, F.A., Chakraborty, J., Pandini, R.V., Saraiva, S., Nahas, C.S.R., Nahas, S.C.,Nomura,C.H.,2023. Mri-basedradiomicscoreincreasedmrtrg accuracy in predicting rectal cancer response to neoadjuvant therapy. Abdominal Radiology 48, 1911–1920

    2023

  27. [27]

    arXiv preprint arXiv:2010.15327 , year=

    Nguyen, T., Raghu, M., Kornblith, S., 2020. Do wide and deep networks learn the same things? uncovering how neural net- work representations vary with width and depth. arXiv preprint arXiv:2010.15327

    work page arXiv 2020

  28. [28]

    Clinically applicable segmentation of head and neck anatomy for radiotherapy: deep learning algorithm development and validation study

    Nikolov, S., Blackwell, S., Zverovitch, A., Mendes, R., Livne, M., De Fauw, J., Patel, Y., Meyer, C., Askham, H., Romera-Paredes, B., et al., 2021. Clinically applicable segmentation of head and neck anatomy for radiotherapy: deep learning algorithm development and validation study. Journal of medical Internet research 23, e26151

    2021

  29. [29]

    Pytorch:Animperativestyle,high-performancedeeplearninglibrary

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al., 2019. Pytorch:Animperativestyle,high-performancedeeplearninglibrary. Advances in neural information processing systems

    2019

  30. [30]

    Clinical utility of radiomics at baseline rectal mri to predict complete response of rectal cancer after chemoradiation therapy

    Petkovska, I., Tixier, F., Ortiz, E.J., Golia Pernicka, J.S., Paroder, V., Bates, D.D., Horvat, N., Fuqua, J., Schilsky, J., Gollub, M.J., et al., 2020. Clinical utility of radiomics at baseline rectal mri to predict complete response of rectal cancer after chemoradiation therapy. Abdominal Radiology 45, 3608–3617

    2020

  31. [31]

    Transfusion: Understanding transfer learning for medical imaging

    Raghu, M., Zhang, C., Kleinberg, J., Bengio, S., 2019. Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems 32

    2019

  32. [32]

    U-net: Convolutional networks for biomedical image segmentation, in: International Con- ference on Medical Image Computing and Computer-Assisted Inter- vention, Springer

    Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation, in: International Con- ference on Medical Image Computing and Computer-Assisted Inter- vention, Springer

    2015

  33. [33]

    Cancer statistics, 2022

    Siegel, R.L., Miller, K.D., Fuchs, H.E., Jemal, A., 2022. Cancer statistics, 2022. CA: a cancer journal for clinicians 72, 7–33

    2022

  34. [34]

    Leading cancer deaths in people younger than 50 years

    Siegel, R.L., Wagle, N.S., Jemal, A., 2026. Leading cancer deaths in people younger than 50 years. JAMA

    2026

  35. [35]

    Feature selection via dependence maximization

    Song, L., Smola, A., Gretton, A., Bedo, J., Borgwardt, K., 2012. Feature selection via dependence maximization. Journal of Machine Learning Research

    2012

  36. [36]

    Self-supervised pre-training of swin transformers for 3d medical image analysis, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition

    Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A., 2022. Self-supervised pre-training of swin transformers for 3d medical image analysis, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition

    2022

  37. [37]

    Reliability of tumor segmentation in glioblastoma: Impact on the robustness of mri-radiomic features

    Tixier, F., Um, H., Young, R.J., Veeraraghavan, H., 2019. Reliability of tumor segmentation in glioblastoma: Impact on the robustness of mri-radiomic features. Medical Physics 46, 3582–3591. URL: https://pubmed.ncbi.nlm.nih.gov/31131906/, doi:10.1002/mp.13624

    work page doi:10.1002/mp.13624 2019

  38. [38]

    Trebeschi, S., van Griethuysen, J.J., Lambregts, D.M., Lahaye, M.J., Parmar, C., Bakers, F.C., Peters, N.H., Beets-Tan, R.G., Aerts, H.J.,

  39. [39]

    Scientificreports7,5301

    Deep learning for fully-automated localization and segmenta- tionofrectalcanceronmultiparametricmr. Scientificreports7,5301

  40. [40]

    Evaluation of measures for assessing time-saving of automatic organ-at-risk segmentation in radiotherapy

    Vaassen,F.,Hazelaar,C.,Vaniqui,A.,Gooding,M.,VanderHeyden, B., Canters, R., Van Elmpt, W., 2020. Evaluation of measures for assessing time-saving of automatic organ-at-risk segmentation in radiotherapy. Physics and Imaging in Radiation Oncology 13, 1–6

    2020

  41. [41]

    The accuracy of multi- detectorrowctfortheassessmentoftumorinvasionofthemesorectal fascia in primary rectal cancer

    Vliegen,R.,Dresen,R.,Beets,G.,Daniels-Gooszen,A.,Kessels,A., van Engelshoven, J., Beets-Tan, R., 2008. The accuracy of multi- detectorrowctfortheassessmentoftumorinvasionofthemesorectal fascia in primary rectal cancer. Abdominal imaging 33, 604–610

    2008

  42. [42]

    Triad:Visionfoundationmodelfor3dmagnetic resonance imaging

    Wang, S., Safari, M., Li, Q., Chang, C.W., Qiu, R.L., Roper, J., Yu, D.S.,Yang,X.,2025. Triad:Visionfoundationmodelfor3dmagnetic resonance imaging. Research Square

    2025

  43. [43]

    Efficient streaming language models with attention sinks, in: International Conference on Learning Representations

    Xiao, G., Tian, Y., Chen, B., Han, S., Lewis, M., 2024. Efficient streaming language models with attention sinks, in: International Conference on Learning Representations

    2024

  44. [44]

    21-gene recurrence score and survival outcomes in the phase iii multicenter tailorx clinical trial

    Yang, S.X., Yu, J., Wang, M., 2024. 21-gene recurrence score and survival outcomes in the phase iii multicenter tailorx clinical trial. JournaloftheNationalComprehensiveCancerNetwork22,376–381. A. Rangnekar et al.:Preprint submitted to ElsevierPage 13 of 13

    2024