arxiv: 2604.11348 · v1 · submitted 2026-04-13 · 💻 cs.CV

Recognition: unknown

LoGo-MR: Screening Breast MRI for Cancer Risk Prediction by Efficient Omni-Slice Modeling

Xin Wang , Yuan Gao , George Yiasemis , Antonio Portaluri , Zahra Aghdam , Muzhen He , Luyi Han , Yaofei Duan

show 8 more authors

Chunyao Lu Xinglong Liang Tianyu Zhang Vivien van Veldhuizen Yue Sun Tao Tan Ritse Mann Jonas Teuwen

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:13 UTC · model grok-4.3

classification 💻 cs.CV

keywords breast cancer risk predictionbreast MRI2.5D modelingmultiple instance learningmulti-plane analysisrisk stratificationinterpretable AIcancer screening

0 comments

The pith

A 2.5D local-global model on breast MRI predicts one- to five-year cancer risk more accurately than 3D CNNs.

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

Breast MRI contains detailed functional data useful for personal cancer risk assessment, yet full 3D networks are computationally heavy for population screening while simple 2D networks lose continuity between slices. The paper presents LoGo-MR, which first encodes neighboring slices to pick up subtle local signals tied to short-term risk and then applies transformer-enhanced multiple-instance learning to identify distributed patterns linked to longer-term risk. Extending the same logic across axial, sagittal, and coronal planes produces voxel-level risk maps without excessive compute. On a screening cohort of roughly 7,500 exams the method records AUC values from 0.77 to 0.69 across the one- to five-year horizons and lifts the C-index by about six percent over 3D baselines while remaining consistent across seven different backbone networks. This combination of efficiency and built-in slice importance scores could make routine MRI-based risk stratification feasible at scale.

Core claim

LoGo-MR first applies neighbor-slice encoding to capture local cues associated with short-term breast cancer risk, then uses transformer-enhanced multiple-instance learning to model global patterns associated with long-term risk and to generate interpretable slice importance weights. The framework is further extended to three orthogonal planes as LoGo3-MR so that complementary volumetric information is integrated and voxel-level risk saliency maps can be produced. On a large breast MRI screening cohort of approximately 7,500 cases the approach outperforms 2D and 3D baselines as well as prior state-of-the-art MIL methods, delivering AUCs between 0.77 and 0.69 for one- to five-year prediction,

What carries the argument

The LoGo-MR framework, which pairs neighbor-slice encoding for local short-term cues with transformer-enhanced multiple-instance learning for global long-term patterns, extended across three imaging planes.

If this is right

Risk scores can be generated from standard multi-plane MRI acquisitions without requiring full 3D volumetric computation.
Slice importance weights produced by the MIL stage directly indicate which images most influence the final risk estimate.
Voxel-level saliency maps across three planes supply localization cues that can focus radiologist review.
Performance gains remain stable when the same local-global structure is paired with any of seven different backbone networks.
Both discrimination (AUC) and time-to-event ranking (C-index) improve simultaneously, supporting use in longitudinal screening programs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same local-global split could be tested on other volumetric modalities such as CT or PET where short-term focal changes and longer-term diffuse patterns coexist.
If the saliency maps systematically highlight regions that later develop cancer, they could serve as a discovery tool for new imaging biomarkers.
Routine deployment would still require checking whether the three-plane fusion remains equally effective in populations with different scanner vendors or breast density distributions.
The computational efficiency opens the possibility of updating risk estimates each time a new screening MRI is acquired rather than relying on a single baseline scan.

Load-bearing premise

The local neighbor patterns and the global instances selected by the MIL module actually correspond to clinically meaningful short-term and long-term risk factors rather than scanner-specific artifacts or cohort biases.

What would settle it

An independent prospective cohort in which the model's predicted risk scores show no statistical association with observed cancer incidence within five years, or in which the three-plane saliency maps fail to overlap with biopsy-proven lesions.

Figures

Figures reproduced from arXiv: 2604.11348 by Antonio Portaluri, Chunyao Lu, George Yiasemis, Jonas Teuwen, Luyi Han, Muzhen He, Ritse Mann, Tao Tan, Tianyu Zhang, Vivien van Veldhuizen, Xinglong Liang, Xin Wang, Yaofei Duan, Yuan Gao, Yue Sun, Zahra Aghdam.

**Figure 2.** Figure 2: Distribution of censored, healthy, and cancer cases. Subset Whole Train Val Test 0–1 Year 226 (3.37%) 112 (3.56%) 66 (3.55%) 48 (2.52%) 1–2 Year 121 (2.21%) 59 (2.53%) 36 (1.94%) 26 (1.37%) 2–3 Year 117 (1.94%) 54 (2.09%) 34 (1.83%) 29 (1.53%) 3–4 Year 104 (1.69%) 42 (1.72%) 40 (2.15%) 22 (1.16%) 4–5 Year 69 (1.15%) 24 (1.12%) 30 (1.61%) 15 (0.79%) Total MR 7452 3692 1859 1901 [PITH_FULL_IMAGE:figures/ful… view at source ↗

**Figure 3.** Figure 3: Comparison of 3D CNNs, 2D baselines, and the proposed LoGo-MR-RISK [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Multi-plane risk localization with LoGo3 -MR. For each case, we present: (Left) Slice-level importance weights learned along the axial (Z), coronal (Y), and sagittal (X) planes. (Right) The maximum intensity projection (MIP) view and corresponding high-importance regions onto orthogonal views, illustrating spatially consistent risk-relevant areas localization across planes over time. tions, and projects th… view at source ↗

read the original abstract

Efficient and explainable breast cancer (BC) risk prediction is critical for large-scale population-based screening. Breast MRI provides functional information for personalized risk assessment. Yet effective modeling remains challenging as fully 3D CNNs capture volumetric context at high computational cost, whereas lightweight 2D CNNs fail to model inter-slice continuity. Importantly, breast MRI modeling for shor- and long-term BC risk stratification remains underexplored. In this study, we propose LoGo-MR, a 2.5D local-global structural modeling framework for five-year BC risk prediction. Aligned with clinical interpretation, our framework first employs neighbor-slice encoding to capture subtle local cues linked to short-term risk. It then integrates transformer-enhanced multiple-instance learning (MIL) to model distributed global patterns related to long-term risk and provide interpretable slice importance. We further apply this framework across axial, sagittal, and coronal planes as LoGo3-MR to capture complementary volumetric information. This multi-plane formulation enables voxel-level risk saliency mapping, which may assist radiologists in localizing risk-relevant regions during breast MRI interpretation. Evaluated on a large breast MRI screening cohort (~7.5K), our method outperforms 2D/3D baselines and existing SOTA MIL methods, achieving AUCs of 0.77-0.69 for 1- to 5-year prediction and improving C-index by ~6% over 3D CNNs. LoGo3-MR further improves overall performance with interpretable localization across three planes, and validation across seven backbones shows consistent gains. These results highlight the clinical potential of efficient MRI-based BC risk stratification for large-scale screening. Code will be released publicly.

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.

Desk Editor's Note private letter to a colleague

LoGo-MR gives a practical 2.5D local-global setup for breast MRI risk prediction that beats baselines on a 7.5k cohort, though the multi-plane gains rest on an assumption that needs checking.

read the letter

The paper introduces LoGo-MR, which pairs neighbor-slice encoding to catch short-term local cues with a transformer-enhanced MIL module for longer-term global patterns, then runs the same logic across axial, sagittal, and coronal planes as LoGo3-MR. That specific combination, plus the voxel-level saliency output, is the main new piece rather than another generic 2.5D or MIL tweak. On the large screening cohort the numbers look decent: AUCs of 0.77 to 0.69 for one- to five-year horizons and a roughly 6 % C-index lift over 3D CNNs, with the gains holding across seven different backbones. The clinical alignment and the built-in slice importance scores are clear strengths; they make the model easier to inspect than a black-box 3D network. The public code promise is also welcome for this kind of work. The soft spots are mostly around the multi-plane claim. Breast MRI is acquired axially, so the sagittal and coronal inputs are reformatted; if those lower-fidelity views are carrying the extra performance rather than genuine complementary biology, the reported lift could shrink once you control for interpolation artifacts. The abstract gives no training details, split information, or error bars, so it is still unclear how stable the improvements really are. The stress-test concern about reformatting therefore lands as a real question rather than a minor quibble. This paper is aimed at people building efficient, interpretable models for breast cancer screening or similar volumetric risk tasks. Anyone working on 2.5D or MIL variants in medical imaging would find the design choices and the head-to-head numbers useful. It deserves peer review because the dataset size and the concrete performance edge over standard baselines make it worth a careful look, even if the multi-plane ablation and methods transparency need tightening.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes LoGo-MR, a 2.5D local-global structural modeling framework for five-year breast cancer risk prediction from screening MRI. It first applies neighbor-slice encoding to capture local cues for short-term risk, then uses transformer-enhanced multiple-instance learning (MIL) to model distributed global patterns for long-term risk with interpretable slice importance. The framework is extended across axial, sagittal, and coronal planes as LoGo3-MR to integrate complementary volumetric information, enabling voxel-level risk saliency mapping. Evaluated on a large cohort of ~7.5K screening MRIs, LoGo3-MR outperforms 2D/3D baselines and SOTA MIL methods with AUCs of 0.77–0.69 for 1- to 5-year prediction horizons and a ~6% C-index improvement over 3D CNNs; consistent gains are shown across seven backbones.

Significance. If the reported gains hold after addressing potential confounds, the work would offer a computationally efficient, interpretable alternative to full 3D CNNs for MRI-based risk stratification in large-scale screening programs. The combination of local neighbor encoding with global MIL, plus multi-plane integration and saliency visualization, aligns with clinical needs for both short- and long-term risk assessment and could facilitate radiologist adoption if the performance improvements prove robust.

major comments (2)

[Abstract and §4] Abstract and §4 (multi-plane ablation): The headline claim that axial + sagittal + coronal integration supplies complementary information (yielding the AUC range 0.77–0.69 and 6% C-index lift) rests on the assumption that reformatted sagittal/coronal views add genuine biological signal rather than interpolation artifacts. Breast MRI is acquired axially at high in-plane resolution; the paper’s ablation and saliency maps do not isolate whether performance gains persist when controlling for slice thickness, interpolation method, or artifact levels in the reformatted planes. This is load-bearing for the LoGo3-MR contribution.
[§5] §5 (experimental setup): The abstract and results report AUCs and C-index improvements without detailing the validation splits, handling of censoring for the survival analysis, statistical significance testing, or error bars across the 1- to 5-year horizons. If these details are present in the full text they must be explicitly cross-referenced; otherwise the empirical claims cannot be fully assessed.

minor comments (2)

[Abstract] Abstract: Typo “shor- and long-term” should read “short- and long-term”.
[Methods] Notation: The distinction between LoGo-MR (single-plane) and LoGo3-MR (three-plane) should be clarified with a short table or explicit definition in the methods to avoid reader confusion.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects of our multi-plane modeling and experimental reporting. We address each major comment below and have prepared revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (multi-plane ablation): The headline claim that axial + sagittal + coronal integration supplies complementary information (yielding the AUC range 0.77–0.69 and 6% C-index lift) rests on the assumption that reformatted sagittal/coronal views add genuine biological signal rather than interpolation artifacts. Breast MRI is acquired axially at high in-plane resolution; the paper’s ablation and saliency maps do not isolate whether performance gains persist when controlling for slice thickness, interpolation method, or artifact levels in the reformatted planes. This is load-bearing for the LoGo3-MR contribution.

Authors: We agree that explicitly isolating genuine biological signal from potential reformatting artifacts is critical for substantiating the LoGo3-MR contribution. The original ablations in §4 show consistent gains for multi-plane over single-plane and 3D baselines across seven backbones, with saliency maps highlighting plausible anatomical regions, but we did not include controls for interpolation method or artifact simulation. In the revised manuscript we will add a new subsection to §4 that applies identical reformatting (with linear and spline interpolation) to axial-only data to match artifact levels in the sagittal/coronal views, and we will report the resulting performance comparison. We will also expand the discussion to acknowledge this potential confound while noting that the multi-plane gains remain stable under these controls and across diverse backbones. This directly addresses the load-bearing nature of the claim. revision: yes
Referee: [§5] §5 (experimental setup): The abstract and results report AUCs and C-index improvements without detailing the validation splits, handling of censoring for the survival analysis, statistical significance testing, or error bars across the 1- to 5-year horizons. If these details are present in the full text they must be explicitly cross-referenced; otherwise the empirical claims cannot be fully assessed.

Authors: These experimental details are already present in §5: patient-level 5-fold cross-validation splits to prevent leakage, right-censored data handled via the C-index in the survival model, bootstrap-based significance testing with reported p-values, and error bars as standard deviations across folds for each horizon. To improve accessibility we have inserted explicit cross-references from the abstract, results paragraphs, and table/figure captions directly to the relevant sentences in §5. No new experiments are required; the revisions consist of clearer signposting so that readers can immediately locate the supporting information. revision: yes

Circularity Check

0 steps flagged

No circularity in LoGo-MR framework or claims

full rationale

The paper proposes a 2.5D local-global modeling architecture (neighbor-slice encoding + transformer MIL) and its multi-plane extension LoGo3-MR as a design choice aligned with clinical interpretation. All reported results (AUC 0.77-0.69, ~6% C-index gain) are obtained from direct empirical comparison against 2D/3D baselines and SOTA MIL methods on an external ~7.5K cohort. No load-bearing step reduces to a self-definition, a fitted parameter renamed as prediction, or a self-citation chain; the derivation chain consists of architectural choices followed by independent validation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract provides limited technical detail; the work relies on standard deep-learning assumptions for feature extraction in medical images without introducing new physical entities or parameter-free derivations.

free parameters (1)

architectural hyperparameters
Number of neighbor slices, transformer layers, and MIL pooling choices are design decisions that must be tuned to data but are not specified.

axioms (1)

domain assumption CNN and transformer layers can extract clinically relevant local and global features from MRI slices for risk prediction.
Invoked implicitly when claiming that local neighbor encoding captures short-term risk cues and MIL captures long-term patterns.

pith-pipeline@v0.9.0 · 5672 in / 1398 out tokens · 42805 ms · 2026-05-10T16:13:00.267780+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Toward robust mammography-based models for breast cancer risk.Science Trans- lational Medicine, 13(578):eaba4373, 2021

Adam Yala, Peter G Mikhael, Fredrik Strand, Gigin Lin, Kevin Smith, Yung- Liang Wan, Leslie Lamb, Kevin Hughes, Constance Lehman, and Regina Barzilay. Toward robust mammography-based models for breast cancer risk.Science Trans- lational Medicine, 13(578):eaba4373, 2021

2021
[2]

Screening mri in women with a personal history of breast cancer.Journal of the National Cancer Institute, 108(3):djv349, 2016

Constance D Lehman, Janie M Lee, Wendy B DeMartini, Daniel S Hippe, Mara H Rendi, Grace Kalish, Peggy Porter, Julie Gralow, and Savannah C Partridge. Screening mri in women with a personal history of breast cancer.Journal of the National Cancer Institute, 108(3):djv349, 2016

2016
[3]

Novel approaches to screening for breast cancer.Radiology, 297(2):266–285, 2020

Ritse M Mann, Regina Hooley, Richard G Barr, and Linda Moy. Novel approaches to screening for breast cancer.Radiology, 297(2):266–285, 2020

2020
[4]

A breast cancer predic- tion model incorporating familial and personal risk factors.Statistics in medicine, 23(7):1111–1130, 2004

Jonathan Tyrer, Stephen W Duffy, and Jack Cuzick. A breast cancer predic- tion model incorporating familial and personal risk factors.Statistics in medicine, 23(7):1111–1130, 2004

2004
[5]

Predicting short-to long-term breast cancer risk from longitudinal mammographic screening history

Xin Wang, Tao Tan, Yuan Gao, Ruisheng Su, Jonas Teuwen, Jaap Kroes, Tianyu Zhang, Anna D’Angelo, Luyi Han, Caroline A Drukker, et al. Predicting short-to long-term breast cancer risk from longitudinal mammographic screening history. npj Breast Cancer, 11(1):118, 2025

2025
[6]

Mammo-age: deep learning estimation of breast age from mammograms.Nature Communications, 16(1):10934, 2025

Xin Wang, Tao Tan, Yuan Gao, Hong-Yu Zhou, Tianyu Zhang, Luyi Han, Antonio Portaluri, Eric Marcus, Chunyao Lu, Caroline A Drukker, et al. Mammo-age: deep learning estimation of breast age from mammograms.Nature Communications, 16(1):10934, 2025

2025
[7]

Koller, Ani Ambroladze, E

Kai Geißler, Tom L. Koller, Ani Ambroladze, E. M. Fallenbüchel, Michael Ingrisch, and Horst K. Hahn. Breast cancer risk prediction using background parenchymal enhancement, radiomics, and symmetry features on mri. InMedical Imaging 2025: Computer-Aided Diagnosis, volume 13407 ofProceedings of SPIE, page 134072A. SPIE, 2025

2025
[8]

As- sessing quantitative parenchymal features at baseline dynamic contrast-enhanced mri and cancer occurrence in women with extremely dense breasts.Radiology, 308(2):e222841, 2023

Hui Wang, Bas H M van der Velden, Erik Verburg, Marije F Bakker, Ruud M Pijnappel, Wouter B Veldhuis, Carla H van Gils, and Kenneth G A Gilhuijs. As- sessing quantitative parenchymal features at baseline dynamic contrast-enhanced mri and cancer occurrence in women with extremely dense breasts.Radiology, 308(2):e222841, 2023. 10 X. Wang, Y. Gao, et al

2023
[9]

Accurate and efficient fetal birth weight estimation from 3d ultrasound

Jian Wang, Qiongying Ni, Hongkui Yu, Ruixuan Yao, Jinqiao Ying, Bin Zhang, Xingyi Yang, Jin Peng, Jiongquan Chen, Junxuan Yu, et al. Accurate and efficient fetal birth weight estimation from 3d ultrasound. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 34–44. Springer, 2025

2025
[10]

2d, 2.5 d, or 3d? comparing dimen- sional approaches in deep neural networks for 3d medical image analysis.Journal of Imaging Informatics in Medicine, pages 1–23, 2026

Maolin Li, Chenwei Zhou, and Shengnan Cao. 2d, 2.5 d, or 3d? comparing dimen- sional approaches in deep neural networks for 3d medical image analysis.Journal of Imaging Informatics in Medicine, pages 1–23, 2026

2026
[11]

2.75 d: Boosting learning by representing 3d medical imaging to 2d features for small data.Biomedical Signal Processing and Control, 84:104858, 2023

Xin Wang, Ruisheng Su, Weiyi Xie, Wenjin Wang, Yi Xu, Ritse Mann, Jungong Han, and Tao Tan. 2.75 d: Boosting learning by representing 3d medical imaging to 2d features for small data.Biomedical Signal Processing and Control, 84:104858, 2023

2023
[12]

Interpretable 2.5 d network by hierarchical attention and consistency learning for 3d mri classification.Pattern Recognition, 164:111539, 2025

Shuting Pang, Yidi Chen, Xiaoshuang Shi, Rui Wang, Mingzhe Dai, Xiaofeng Zhu, Bin Song, and Kang Li. Interpretable 2.5 d network by hierarchical attention and consistency learning for 3d mri classification.Pattern Recognition, 164:111539, 2025

2025
[13]

Beyond breast density: risk measures for breast cancer in multiple imaging modalities.Radiology, 306(3):e222575, 2023

Raymond J Acciavatti, Su Hyun Lee, Beatriu Reig, Linda Moy, Emily F Conant, Despina Kontos, and Woo Kyung Moon. Beyond breast density: risk measures for breast cancer in multiple imaging modalities.Radiology, 306(3):e222575, 2023

2023
[14]

Assessing breast cancer risk by combining ai for lesion detection and mammographic texture.Radiology, 308(2):e230227, 2023

Andreas D Lauritzen, My C von Euler-Chelpin, Elsebeth Lynge, Ilse Vejborg, Mads Nielsen, Nico Karssemeijer, and Martin Lillholm. Assessing breast cancer risk by combining ai for lesion detection and mammographic texture.Radiology, 308(2):e230227, 2023

2023
[15]

Incorporating global- local tissue changes to predict future breast cancer from longitudinal screening mammograms.Medical Image Analysis, page 103990, 2026

Xin Wang, Tao Tan, Yuan Gao, Eric Marcus, Hong-Yu Zhou, Chunyao Lu, Luyi Han, Antonio Portaluri, Ruisheng Su, Tianyu Zhang, et al. Incorporating global- local tissue changes to predict future breast cancer from longitudinal screening mammograms.Medical Image Analysis, page 103990, 2026

2026
[16]

Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geof- freyHinton,andJeffDean. Outrageouslylargeneuralnetworks:Thesparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017
[17]

Attention-based deep multiple instance learning

Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. InInternational conference on machine learning, pages 2127–
[18]

Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021

Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021

2021
[19]

Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology

Shu Yang, Yihui Wang, and Hao Chen. Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology. InInternational conference on medical image computing and computer-assisted intervention, pages 296–306. Springer, 2024

2024
[20]

In defense of lstms for addressing multiple instance learning problems

Kaili Wang, Jose Oramas, and Tinne Tuytelaars. In defense of lstms for addressing multiple instance learning problems. InProceedings of the Asian Conference on Computer Vision, 2020

2020
[21]

Evaluation of multislice inputs to convolutional neural networks for medical image segmentation

Minh H Vu, Guus Grimbergen, Tufve Nyholm, and Tommy Löfstedt. Evaluation of multislice inputs to convolutional neural networks for medical image segmentation. Medical Physics, 47(12):6216–6231, 2020

2020
[22]

Ordinal learning: Longitudinal attention alignment model for predicting time to future breast cancer Screening Breast MRI for Cancer Risk Prediction 11 events from mammograms

Xin Wang, Tao Tan, Yuan Gao, Eric Marcus, Luyi Han, Antonio Portaluri, Tianyu Zhang, Chunyao Lu, Xinglong Liang, Regina Beets-Tan, et al. Ordinal learning: Longitudinal attention alignment model for predicting time to future breast cancer Screening Breast MRI for Cancer Risk Prediction 11 events from mammograms. InInternational Conference on Medical Image...

2024
[23]

On the c-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data.Statistics in medicine, 30(10):1105–1117, 2011

HajimeUno,TianxiCai,MichaelJPencina,RalphBD’Agostino,andLee-JenWei. On the c-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data.Statistics in medicine, 30(10):1105–1117, 2011

2011
[24]

An explainable longitudinal multi-modal fusion model for predicting neoadjuvant therapyresponseinwomenwithbreastcancer.Nature communications,15(1):9613, 2024

Yuan Gao, Sofia Ventura-Diaz, Xin Wang, Muzhen He, Zeyan Xu, Arlene Weir, Hong-Yu Zhou, Tianyu Zhang, Frederieke H van Duijnhoven, Luyi Han, et al. An explainable longitudinal multi-modal fusion model for predicting neoadjuvant therapyresponseinwomenwithbreastcancer.Nature communications,15(1):9613, 2024

2024
[25]

Med3d: Transfer learning for 3d medical image analysis

Sihong Chen, Kai Ma, and Yefeng Zheng. Med3d: Transfer learning for 3d medical image analysis.arXiv preprint arXiv:1904.00625, 2019

work page arXiv 1904