A Dynamic Prognostic Prediction Method for Colorectal Cancer Liver Metastasis

Chengchang Pan; Honggang Qi; Wei Yang; Yan su; Yiran Zhu; Zesheng Li

arxiv: 2505.03123 · v3 · submitted 2025-05-06 · 📡 eess.IV · cs.CV· cs.MM

A Dynamic Prognostic Prediction Method for Colorectal Cancer Liver Metastasis

Wei Yang , Yiran Zhu , Yan su , Zesheng Li , Chengchang Pan , Honggang Qi This is my paper

Pith reviewed 2026-05-22 17:23 UTC · model grok-4.3

classification 📡 eess.IV cs.CVcs.MM

keywords colorectal cancerliver metastasisprognostic predictiondynamic modelingdeep learningsurvival analysisresidual evolutionlongitudinal trajectories

0 comments

The pith

DyPro generates 12-step postoperative trajectory sequences from an initial patient representation to predict survival in colorectal cancer liver metastasis.

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

The paper introduces DyPro, a framework that moves beyond static single-snapshot models by evolving an initial patient representation through autoregressive residual updates. It produces a fixed-length sequence of latent trajectory snapshots that together encode longitudinal disease progression, tumor distribution, and multimodal clinical data. These integrated dynamics are then used to forecast overall survival and disease-free survival. On the MSKCC CRLM dataset the approach yields a C-index of 0.755 for OS and 0.714 for DFS under repeated cross-validation. The authors position the method as a source of quantitative risk information that could guide adjuvant therapy and follow-up decisions.

Core claim

DyPro infers postoperative latent trajectories via residual dynamic evolution: starting from a single initial patient representation, it generates a 12-step sequence of trajectory snapshots through autoregressive residual updates and integrates the sequence to predict recurrence and survival outcomes.

What carries the argument

Autoregressive residual updates that evolve an initial patient representation into a 12-step sequence of latent trajectory snapshots.

If this is right

The framework produces a C-index of 0.755 for overall survival and 0.714 for disease-free survival on the MSKCC CRLM dataset under repeated stratified 5-fold cross-validation.
It yields an OS AUC at one year of 0.920 and an integrated Brier score of 0.143.
The resulting risk scores supply quantitative cues that can inform adjuvant therapy planning and follow-up scheduling.

Where Pith is reading between the lines

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

If the residual-update mechanism proves stable across institutions, the same architecture could be retrained on other metastatic cancers that exhibit high recurrence heterogeneity.
The fixed 12-step horizon suggests a natural test: whether shortening or lengthening the generated sequence improves calibration on datasets with different follow-up lengths.

Load-bearing premise

The model assumes that successive residual updates applied to one starting representation can faithfully reproduce real longitudinal disease dynamics and multimodal clinical information without introducing artifacts or omitting key biological factors.

What would settle it

A prospective cohort study that records actual serial imaging, biomarker, and clinical measurements after surgery and checks whether the model's generated 12-step snapshots align with the observed changes in those measurements.

Figures

Figures reproduced from arXiv: 2505.03123 by Chengchang Pan, Honggang Qi, Wei Yang, Yan su, Yiran Zhu, Zesheng Li.

**Figure 1.** Figure 1: Overview of DyPro. Preoperative CT and clinical data are mapped to a heterogeneous patient graph, evolved by latent residual dynamics, and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: Visualization of the dataset, including segmentation of the liver, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

read the original abstract

Colorectal cancer liver metastasis (CRLM) exhibits high postoperative recurrence and pronounced prognostic heterogeneity, challenging individualized management. Existing prognostic approaches often rely on static representations from a single postoperative snapshot, and fail to jointly capture tumor spatial distribution, longitudinal disease dynamics, and multimodal clinical information, limiting predictive accuracy. We propose DyPro, a deep learning framework that infers postoperative latent trajectories via residual dynamic evolution. Starting from an initial patient representation, DyPro generates a 12-step sequence of trajectory snapshots through autoregressive residual updates and integrates them to predict recurrence and survival outcomes. On the MSKCC CRLM dataset, DyPro achieves strong discrimination under repeated stratified 5-fold cross-validation, reaching a C-index of 0.755 for OS and 0.714 for DFS, with OS AUC@1y of 0.920 and OS IBS of 0.143. DyPro provides quantitative risk cues to support adjuvant therapy planning and follow-up scheduling.

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

DyPro generates 12-step trajectories via autoregressive residuals for CRLM prognosis and reports decent C-index numbers, but the validation is thin on baselines and checks for plausible dynamics.

read the letter

I looked at the DyPro paper on dynamic prognosis for colorectal cancer liver metastasis. The main takeaway is that they start from one postoperative patient representation and use residual updates to build an explicit 12-step latent trajectory sequence, then feed the integrated features into outcome prediction for overall survival and disease-free survival. That specific setup for jointly handling spatial tumor distribution, longitudinal evolution, and multimodal clinical data is not something I recall from the earlier work they reference. It is a straightforward extension of time-series deep learning ideas into this high-recurrence setting, and the reported numbers on the MSKCC dataset under repeated stratified 5-fold cross-validation are respectable: C-index 0.755 for OS, 0.714 for DFS, AUC@1y 0.920, and IBS 0.143. Those figures suggest the approach could give useful risk cues for adjuvant planning. The paper does a clean job of stating the clinical motivation and describing how the autoregressive residual operator produces the snapshots without obvious circularity in the abstract. The central assumption, though, is that those generated trajectories faithfully reflect real disease dynamics rather than artifacts from the update rule. The dataset provides mostly static or limited follow-up labels, and the abstract does not describe direct comparisons to held-out longitudinal sequences or external biological checks. Without baselines, ablations, or error bars, it is hard to tell how much the dynamic component actually adds over simpler static or non-autoregressive models. If the full manuscript includes those controls and shows the trajectories align with observed progression patterns, the claim strengthens considerably. Otherwise the performance numbers stand mostly on their own. This is the sort of paper that would interest researchers working on prognostic models in GI oncology or on dynamic deep learning for medical time series. A reader who wants concrete examples of residual evolution applied to multimodal cancer data could extract some practical ideas from it. It is solid enough on the surface to warrant sending out for peer review so the methodological gaps can be examined and the results tested more thoroughly.

Referee Report

2 major / 2 minor

Summary. The paper proposes DyPro, a deep learning framework for prognostic prediction in colorectal cancer liver metastasis (CRLM). It infers postoperative latent trajectories via residual dynamic evolution: starting from a single initial patient representation, it generates a 12-step sequence of trajectory snapshots through autoregressive residual updates that integrate tumor spatial distribution, longitudinal dynamics, and multimodal clinical information. These integrated features are then used to predict recurrence and survival outcomes. On the MSKCC CRLM dataset, under repeated stratified 5-fold cross-validation, DyPro reports a C-index of 0.755 for overall survival (OS), 0.714 for disease-free survival (DFS), OS AUC@1y of 0.920, and OS IBS of 0.143, positioning the method as a tool for quantitative risk assessment to guide adjuvant therapy and follow-up.

Significance. If the generated trajectories prove to be biologically plausible rather than artifacts, DyPro could meaningfully advance individualized CRLM management by moving beyond static single-snapshot representations to dynamic, multimodal predictions. The reported discrimination metrics (C-index and AUC) indicate potential utility for risk stratification, but the significance hinges on whether the autoregressive process faithfully captures real longitudinal disease evolution without introducing unvalidated artifacts, especially given the dataset's limited longitudinal detail.

major comments (2)

[Methods (trajectory generation) and Results (performance evaluation)] The central performance claims rest on the assumption that autoregressive residual updates initialized from one postoperative snapshot can faithfully model real longitudinal dynamics and multimodal information. The MSKCC CRLM dataset supplies only static or limited follow-up labels, yet no validation of the generated 12-step trajectories against held-out longitudinal sequences or external biological priors is described; this is load-bearing for interpreting the C-index/AUC/IBS values as evidence of improved prognostic accuracy.
[Results and Experiments] The abstract and results report specific metrics (C-index 0.755 for OS, 0.714 for DFS, AUC@1y 0.920, IBS 0.143) under repeated stratified 5-fold CV but provide no baseline comparisons, ablation studies, error bars, or derivation details for the residual evolution operator. Without these, it is unclear whether the reported gains are attributable to the dynamic trajectory component or to other factors.

minor comments (2)

[Methods] Clarify the exact number of patients and key hyperparameters (e.g., number of trajectory steps, residual update formulation) in the methods section for reproducibility.
[Abstract] The abstract would benefit from a brief statement on how the integrated trajectory features are aggregated for the final prediction head.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which help clarify the strengths and limitations of our DyPro framework for dynamic prognostic prediction in colorectal cancer liver metastasis. We address each major comment below with clarifications from the manuscript and indicate planned revisions to enhance transparency and rigor.

read point-by-point responses

Referee: [Methods (trajectory generation) and Results (performance evaluation)] The central performance claims rest on the assumption that autoregressive residual updates initialized from one postoperative snapshot can faithfully model real longitudinal dynamics and multimodal information. The MSKCC CRLM dataset supplies only static or limited follow-up labels, yet no validation of the generated 12-step trajectories against held-out longitudinal sequences or external biological priors is described; this is load-bearing for interpreting the C-index/AUC/IBS values as evidence of improved prognostic accuracy.

Authors: The MSKCC CRLM dataset indeed consists primarily of static postoperative representations and outcome labels with limited dense longitudinal sequences available for direct comparison. DyPro's residual dynamic evolution is formulated to generate latent trajectories that integrate tumor spatial distribution, multimodal clinical features, and autoregressive updates to support outcome prediction, rather than to reconstruct observed time-series data. While we did not perform explicit validation of the 12-step trajectories against held-out real longitudinal sequences (due to data constraints in the cohort), the reported discrimination metrics arise from end-to-end training on observed survival and recurrence endpoints. In the revision, we will add a limitations subsection discussing this point, include qualitative visualizations of sample trajectories, and report sensitivity analyses to initialization to better contextualize the results. revision: yes
Referee: [Results and Experiments] The abstract and results report specific metrics (C-index 0.755 for OS, 0.714 for DFS, AUC@1y 0.920, IBS 0.143) under repeated stratified 5-fold CV but provide no baseline comparisons, ablation studies, error bars, or derivation details for the residual evolution operator. Without these, it is unclear whether the reported gains are attributable to the dynamic trajectory component or to other factors.

Authors: The full manuscript provides the mathematical formulation of the residual evolution operator in the Methods section, including the autoregressive update equations that combine spatial and multimodal embeddings. However, we agree that explicit baseline comparisons, ablation studies isolating the dynamic component, and error bars from the repeated stratified 5-fold cross-validation are not sufficiently detailed in the current results presentation. In the revised version, we will expand the Experiments section to include (i) comparisons against static multimodal baselines and standard survival models, (ii) ablation variants without the residual dynamic evolution, and (iii) mean and standard deviation metrics across the repeated folds to quantify variability. revision: yes

Circularity Check

0 steps flagged

No significant circularity: trajectory generation and outcome integration form an independent modeling chain

full rationale

The paper describes a standard deep learning pipeline in which an initial patient representation is evolved via autoregressive residual updates to produce a 12-step latent trajectory, after which the integrated features are used to predict survival and recurrence outcomes. This structure is trained and evaluated under repeated stratified 5-fold cross-validation on the MSKCC CRLM dataset, with reported metrics (C-index, AUC, IBS) serving as external performance measures rather than definitional inputs. No equation or step reduces the generated trajectories or final predictions to the target labels by construction, nor does any load-bearing claim rest on a self-citation that itself lacks independent verification. The derivation therefore remains self-contained against the held-out validation folds.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard deep learning assumptions plus the domain-specific premise that latent trajectories can be autoregressively evolved from a single postoperative snapshot.

free parameters (1)

number of trajectory steps
Fixed at 12 steps in the described sequence generation process.

axioms (1)

domain assumption An initial patient representation can be evolved autoregressively via residual updates to capture longitudinal disease dynamics.
Invoked in the description of generating the 12-step sequence of trajectory snapshots.

pith-pipeline@v0.9.0 · 5707 in / 1243 out tokens · 39465 ms · 2026-05-22T17:23:45.444553+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/ArithmeticFromLogic.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

DyPro generates a 12-step sequence of trajectory snapshots through autoregressive residual updates ... H(t) = H(t-1) + F_θ([H(t-1); e_t], E)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

[1]

Multimodal treatment strategies for colorectal liver metastases,

D. L. Birrer, C. Tschuor, C. S. Reineret al., “Multimodal treatment strategies for colorectal liver metastases,”Swiss Medical Weekly, vol. 151, p. w20390, 2021

work page 2021
[2]

Clinical score for predicting recur- rence after hepatic resection for metastatic colorectal cancer: analysis of 1001 consecutive cases,

Y . Fong, J. Fortner, R. L. Sunet al., “Clinical score for predicting recur- rence after hepatic resection for metastatic colorectal cancer: analysis of 1001 consecutive cases,”Annals of Surgery, vol. 230, no. 3, pp. 309– 321, 1999

work page 1999
[3]

Clinical prediction models for prognosis of colorectal liver metastases: a com- prehensive review of regression-based and machine learning models,

S. Kokkinakis, I. A. Ziogas, J. D. Llaque Salazaret al., “Clinical prediction models for prognosis of colorectal liver metastases: a com- prehensive review of regression-based and machine learning models,” Cancers (Basel), vol. 16, no. 9, p. 1645, 2024

work page 2024
[4]

Dynamic radiomics for predicting the efficacy of antiangiogenic therapy in colorectal liver metastases,

Y . Qu, X. Zhai, J. Zhanget al., “Dynamic radiomics for predicting the efficacy of antiangiogenic therapy in colorectal liver metastases,” Frontiers in Oncology, vol. 13, p. 992096, 2023

work page 2023
[5]

Predicting treatment response from longitu- dinal images using multi-task deep learning,

C. Jin, H. Yu, J. Keet al., “Predicting treatment response from longitu- dinal images using multi-task deep learning,”Nature Communications, vol. 12, no. 1, p. 1851, 2021

work page 2021
[6]

Preoperative CT and survival data for patients undergoing resection of colorectal liver metastases,

A. L. Simpson, J. Peoples, J. M. Creasyet al., “Preoperative CT and survival data for patients undergoing resection of colorectal liver metastases,”Scientific Data, vol. 11, no. 1, p. 172, 2024

work page 2024
[7]

Cox’s regression model for counting processes: a large sample study,

P. K. Andersen and R. D. Gill, “Cox’s regression model for counting processes: a large sample study,”Annals of Statistics, vol. 10, no. 4, pp. 1100–1120, 1982

work page 1982
[8]

Deep learning from CT images for survival prediction in colorectal liver metastases,

Y . Liu, Y . Zhang, J. Wanget al., “Deep learning from CT images for survival prediction in colorectal liver metastases,”European Radiology, vol. 32, no. 5, pp. 3421–3430, 2022

work page 2022
[9]

Radiomic profiling of glioblastoma: identifying an imaging predictor of patient survival with improved performance over clinical models,

P. Kickingereder, S. Burth, A. Wicket al., “Radiomic profiling of glioblastoma: identifying an imaging predictor of patient survival with improved performance over clinical models,”Radiology, vol. 281, no. 3, pp. 907–918, 2016

work page 2016
[10]

International evaluation of an AI system for breast cancer screening,

S. M. McKinney, M. Sieniek, V . Godboleet al., “International evaluation of an AI system for breast cancer screening,”Nature, vol. 577, no. 7788, pp. 89–94, 2020

work page 2020
[11]

Deep radiomics-based fusion model for prediction of bevacizumab treatment response and outcome in patients with colorectal cancer liver metastases: a multicentre cohort study,

S. Zhou, D. Sun, W. Maoet al., “Deep radiomics-based fusion model for prediction of bevacizumab treatment response and outcome in patients with colorectal cancer liver metastases: a multicentre cohort study,” EClinicalMedicine, vol. 65, p. 102271, 2023

work page 2023
[12]

Graph neural networks in cancer and oncology research: emerging and future trends,

G. Gogoshin and A. S. Rodin, “Graph neural networks in cancer and oncology research: emerging and future trends,”Cancers (Basel), vol. 15, no. 24, p. 5858, 2023

work page 2023
[13]

Deep multimodal graph-based network for survival prediction from highly multiplexed images and patient variables,

X. Fu, E. Patrick, J. Y . H. Yanget al., “Deep multimodal graph-based network for survival prediction from highly multiplexed images and patient variables,”Computerized Medical Imaging and Graphics, vol. 102, p. 102128, 2023

work page 2023
[14]

Dual-stream multi-dependency graph neural network enables precise cancer survival analysis,

Z. Wang, J. Ma, Q. Gaoet al., “Dual-stream multi-dependency graph neural network enables precise cancer survival analysis,”Medical Image Analysis, vol. 97, p. 103252, 2024

work page 2024
[15]

MMGCN: Multi-modal multi-view graph convolutional networks for cancer prognosis prediction,

P. Yang, W. Chen, and H. Qiu, “MMGCN: Multi-modal multi-view graph convolutional networks for cancer prognosis prediction,”Com- puter Methods and Programs in Biomedicine, vol. 257, p. 108400, 2024

work page 2024
[16]

Spatial temporal graph convolutional networks for skeleton-based action recognition,

S. Yan, Y . Xiong, and D. Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7444–7452

work page 2018
[17]

Computed tomography-based habitat analysis for prognostic stratification in colorectal liver metastases,

C. Zhou, H. Xin, L. Qian, Y . Zhang, J. Wang, and J. Luo, “Computed tomography-based habitat analysis for prognostic stratification in colorectal liver metastases,”Cancer Innovation, vol. 4, no. 2, p. e70000, 2025. [Online]. Available: https://doi.org/10.1002/cai2.70000

work page doi:10.1002/cai2.70000 2025
[18]

Finding reproducible and prognostic radiomic features in variable slice thickness contrast enhanced ct of colorectal liver metastases,

J. J. Peoples, M. Hamghalam, I. James, M. Wasim, N. Gangai, H. C. Kang, X. J. Rong, Y . S. Chun, R. K. G. Do, and A. L. Simpson, “Finding reproducible and prognostic radiomic features in variable slice thickness contrast enhanced ct of colorectal liver metastases,” Machine Learning for Biomedical Imaging, vol. 2, pp. 2326–2357,

work page
[19]

Available: https://melba-journal.org/2024:032

[Online]. Available: https://melba-journal.org/2024:032

work page 2024
[20]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Representations (ICLR), 2017. [Online]. Available: https://openreview.net/forum?id=SJU4ayYgl

work page 2017
[21]

Graph attention networks,

P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” inInternational Conference on Learning Representations (ICLR), 2018. [Online]. Available: https://openreview.net/forum?id=rJXMpikCZ

work page 2018
[22]

Inductive representation learning on large graphs,

W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017, pp. 1025–1035

work page 2017
[23]

Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet?

K. Hara, H. Kataoka, and Y . Satoh, “Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet?” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 6546–6555

work page 2018
[24]

A closer look at spatiotemporal convolutions for action recognition,

D. Tran, H. Wang, L. Torresani, J. Ray, Y . LeCun, and M. Paluri, “A closer look at spatiotemporal convolutions for action recognition,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 6450–6459

work page 2018

[1] [1]

Multimodal treatment strategies for colorectal liver metastases,

D. L. Birrer, C. Tschuor, C. S. Reineret al., “Multimodal treatment strategies for colorectal liver metastases,”Swiss Medical Weekly, vol. 151, p. w20390, 2021

work page 2021

[2] [2]

Clinical score for predicting recur- rence after hepatic resection for metastatic colorectal cancer: analysis of 1001 consecutive cases,

Y . Fong, J. Fortner, R. L. Sunet al., “Clinical score for predicting recur- rence after hepatic resection for metastatic colorectal cancer: analysis of 1001 consecutive cases,”Annals of Surgery, vol. 230, no. 3, pp. 309– 321, 1999

work page 1999

[3] [3]

Clinical prediction models for prognosis of colorectal liver metastases: a com- prehensive review of regression-based and machine learning models,

S. Kokkinakis, I. A. Ziogas, J. D. Llaque Salazaret al., “Clinical prediction models for prognosis of colorectal liver metastases: a com- prehensive review of regression-based and machine learning models,” Cancers (Basel), vol. 16, no. 9, p. 1645, 2024

work page 2024

[4] [4]

Dynamic radiomics for predicting the efficacy of antiangiogenic therapy in colorectal liver metastases,

Y . Qu, X. Zhai, J. Zhanget al., “Dynamic radiomics for predicting the efficacy of antiangiogenic therapy in colorectal liver metastases,” Frontiers in Oncology, vol. 13, p. 992096, 2023

work page 2023

[5] [5]

Predicting treatment response from longitu- dinal images using multi-task deep learning,

C. Jin, H. Yu, J. Keet al., “Predicting treatment response from longitu- dinal images using multi-task deep learning,”Nature Communications, vol. 12, no. 1, p. 1851, 2021

work page 2021

[6] [6]

Preoperative CT and survival data for patients undergoing resection of colorectal liver metastases,

A. L. Simpson, J. Peoples, J. M. Creasyet al., “Preoperative CT and survival data for patients undergoing resection of colorectal liver metastases,”Scientific Data, vol. 11, no. 1, p. 172, 2024

work page 2024

[7] [7]

Cox’s regression model for counting processes: a large sample study,

P. K. Andersen and R. D. Gill, “Cox’s regression model for counting processes: a large sample study,”Annals of Statistics, vol. 10, no. 4, pp. 1100–1120, 1982

work page 1982

[8] [8]

Deep learning from CT images for survival prediction in colorectal liver metastases,

Y . Liu, Y . Zhang, J. Wanget al., “Deep learning from CT images for survival prediction in colorectal liver metastases,”European Radiology, vol. 32, no. 5, pp. 3421–3430, 2022

work page 2022

[9] [9]

Radiomic profiling of glioblastoma: identifying an imaging predictor of patient survival with improved performance over clinical models,

P. Kickingereder, S. Burth, A. Wicket al., “Radiomic profiling of glioblastoma: identifying an imaging predictor of patient survival with improved performance over clinical models,”Radiology, vol. 281, no. 3, pp. 907–918, 2016

work page 2016

[10] [10]

International evaluation of an AI system for breast cancer screening,

S. M. McKinney, M. Sieniek, V . Godboleet al., “International evaluation of an AI system for breast cancer screening,”Nature, vol. 577, no. 7788, pp. 89–94, 2020

work page 2020

[11] [11]

Deep radiomics-based fusion model for prediction of bevacizumab treatment response and outcome in patients with colorectal cancer liver metastases: a multicentre cohort study,

S. Zhou, D. Sun, W. Maoet al., “Deep radiomics-based fusion model for prediction of bevacizumab treatment response and outcome in patients with colorectal cancer liver metastases: a multicentre cohort study,” EClinicalMedicine, vol. 65, p. 102271, 2023

work page 2023

[12] [12]

Graph neural networks in cancer and oncology research: emerging and future trends,

G. Gogoshin and A. S. Rodin, “Graph neural networks in cancer and oncology research: emerging and future trends,”Cancers (Basel), vol. 15, no. 24, p. 5858, 2023

work page 2023

[13] [13]

Deep multimodal graph-based network for survival prediction from highly multiplexed images and patient variables,

X. Fu, E. Patrick, J. Y . H. Yanget al., “Deep multimodal graph-based network for survival prediction from highly multiplexed images and patient variables,”Computerized Medical Imaging and Graphics, vol. 102, p. 102128, 2023

work page 2023

[14] [14]

Dual-stream multi-dependency graph neural network enables precise cancer survival analysis,

Z. Wang, J. Ma, Q. Gaoet al., “Dual-stream multi-dependency graph neural network enables precise cancer survival analysis,”Medical Image Analysis, vol. 97, p. 103252, 2024

work page 2024

[15] [15]

MMGCN: Multi-modal multi-view graph convolutional networks for cancer prognosis prediction,

P. Yang, W. Chen, and H. Qiu, “MMGCN: Multi-modal multi-view graph convolutional networks for cancer prognosis prediction,”Com- puter Methods and Programs in Biomedicine, vol. 257, p. 108400, 2024

work page 2024

[16] [16]

Spatial temporal graph convolutional networks for skeleton-based action recognition,

S. Yan, Y . Xiong, and D. Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7444–7452

work page 2018

[17] [17]

Computed tomography-based habitat analysis for prognostic stratification in colorectal liver metastases,

C. Zhou, H. Xin, L. Qian, Y . Zhang, J. Wang, and J. Luo, “Computed tomography-based habitat analysis for prognostic stratification in colorectal liver metastases,”Cancer Innovation, vol. 4, no. 2, p. e70000, 2025. [Online]. Available: https://doi.org/10.1002/cai2.70000

work page doi:10.1002/cai2.70000 2025

[18] [18]

Finding reproducible and prognostic radiomic features in variable slice thickness contrast enhanced ct of colorectal liver metastases,

J. J. Peoples, M. Hamghalam, I. James, M. Wasim, N. Gangai, H. C. Kang, X. J. Rong, Y . S. Chun, R. K. G. Do, and A. L. Simpson, “Finding reproducible and prognostic radiomic features in variable slice thickness contrast enhanced ct of colorectal liver metastases,” Machine Learning for Biomedical Imaging, vol. 2, pp. 2326–2357,

work page

[19] [19]

Available: https://melba-journal.org/2024:032

[Online]. Available: https://melba-journal.org/2024:032

work page 2024

[20] [20]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Representations (ICLR), 2017. [Online]. Available: https://openreview.net/forum?id=SJU4ayYgl

work page 2017

[21] [21]

Graph attention networks,

P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” inInternational Conference on Learning Representations (ICLR), 2018. [Online]. Available: https://openreview.net/forum?id=rJXMpikCZ

work page 2018

[22] [22]

Inductive representation learning on large graphs,

W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017, pp. 1025–1035

work page 2017

[23] [23]

Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet?

K. Hara, H. Kataoka, and Y . Satoh, “Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet?” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 6546–6555

work page 2018

[24] [24]

A closer look at spatiotemporal convolutions for action recognition,

D. Tran, H. Wang, L. Torresani, J. Ray, Y . LeCun, and M. Paluri, “A closer look at spatiotemporal convolutions for action recognition,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 6450–6459

work page 2018