Click, Predict, Trust: Clinician-Guided and Clinician-Evaluated AI Segmentation for Lung Cancer CT-Based Prognosis within the Knowledge-to-Action Framework

arXiv:2510.17039 · cs.CV · Submitted 2025-10-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Click, Predict, Trust".

Jane: Here's the gist of the paper. Lung cancer is the leading cause of cancer death, with CT central to screening, prognosis, and treatment This study shows that VNet, integrated with SSL, can deliver accurate, reproducible, and clinically trusted tools for CTbased lung cancer prognosis Methodological Framework The research employed a Knowledge-to-Action (KTA) framework to guide the development of a clinician-in-the-loop pipeline to improve reproducibility, prediction, and clinical trust The study analyzed multi-center lung cancer CT data from 999 patients across 12 public datasets Five deep learning models (3D Attention U-Net, ResUNet, VNet, ReconNet, SAM-Med3D) were benchmarked against expert contours on whole and click-point cropped images Segmentation reproducibility was tested using 497 PySERA-extracted radiomics features with various statistical tests Six physicians qualitatively evaluated masks across 21 questions spanning seven domains to assess clinical meaningfulness and workflow integration Model Performance and Stability VNet-based segmentations achieved the best geometric accuracy (Dice 0.83±0.07, IoU 0.71±0.09) and radiomics stability (averaged correlation coefficient=0.76, ICC=0.65) Semi-supervised learning (SSL) consistently outperformed supervised learning (SL) across models The SSL configuration involving LASSO–LGBM–VNet achieved the highest test accuracy (0.88 ± 0.003 accuracy, 0.83 ± 0.004 F1) Clinical and Qualitative Insights Qualitative assessments revealed that radiologists favored VNet for peritumoral representation and smoother boundaries Experts p referred AI to provide initial masks for refinement rather than replace physicians Physicians highlighted that AI masks often introduced noisy, unstable boundaries instead of biologically meaningful differences, limiting prognostic stability Overall assessments indicated no significant differences among the five models in diagnostic or prognostic success based on Friedman tests The study concluded that VNet emerged as the most robust model, achieving superior geometric accuracy and strong radiomics reproducibility The findings demonstrate that SSL pipelines with VNet can enhance prognostic prediction, while clinical reliability requires outcome-linked metrics and physician input Future work includes developing a dedicated clinical software platform integrating VNet-based segmentation with click-point and gaze-tracking functionality The study addressed 10 of 11 barriers to routine adoption of AI in the clinic through its integrated evaluation framework ACKNOWLEDGEMENTS This study was supported by the Virtual Collaboration Group (VirCollab, www.vircollab.com) and the Technological Virtual Collaboration (TECVICO CORP.) based in Vancouver, Canada We gratefully acknowledge funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) Award RGPIN-2023-0357, and Discovery Horizons Grant DH-2025-00119, as well as the UBC Department of Radiology 2023 AI Fund CONFLICT OF INTEREST Drs.

Tom: Mohammad R.

Jane: Salmanpour, Mehdi Maghsudi, Arman Gorji, and Ms.

Tom: Sonya Falahati are affiliated with TECVICO Corp The other co-authors declare no relevant conflicts of interest or disclosures DATA AND CODE AVAILABILITY All code is publicly shared at: https://github.com/MohammadRSalmanpour/Clinician-in-the-Loop-AI-Segmentation-for-CT-BasedPrognosis/tree/main REFERENCE 1.

Paper summary: Jane: So, wrapping up this discussion on "Click, Predict, Trust: Clinician-Guided and Clinician-Evaluated AI Segmentation for Lung Cancer CT-Based Prognosis within the Knowledge-to-Action Framework," what does it really mean for the people listening who aren't deep in the AI weeds?

Tom: It means they’ve shown a concrete pathway to take complex deep learning segmentation models, like VNet, and make them more trustworthy for serious medical use by forcing them through a process that involves both high accuracy testing and direct clinical feedback.

Lu: Essentially, the paper demonstrates that putting an AI tool into a system where it supports the physician's initial mask creation instead of completely replacing it addresses major issues around trust and workflow integration in healthcare.

Meng: For those of us building these things, this study confirms that focusing on reproducibility and stability—like VNet achieved with its zero point eight three dice score—is what actually translates into a reliable tool rather than just a cool algorithm on a laptop.

Jane: Right, so the authors are pointing toward using semi-supervised learning pipelines with models like VNet as the most suitable starting point because they balance high performance with that necessary clinical alignment. The implication is that for AI in diagnostics, success isn't just about getting a high accuracy score; it’s about proving it works consistently when a human expert evaluates the output.

Tom: And they clearly show how to test for those things by using outcome-linked metrics and physician input, which addresses the skepticism barrier head-on. This moves us toward practical implementation rather than just theoretical potential.

Conclusion: Tom: So, we've been looking at how this study tackled that whole pipeline for lung cancer prognosis using VNet and semi-supervised learning to build trust in AI tools, and now we’re hitting the conclusion on this paper by Salmanpour and his team.

Jane: Right. The title itself is pretty clear: Click, Predict, Trust. It tells us the core idea isn't just about getting a good segmentation mask; it's about making sure that mask actually helps the doctor make a decision they trust.

Lu: What’s really interesting here is how they framed it using this Knowledge-to-Action framework. It’s not just building a model and hoping for the best; they built a whole system designed around how doctors actually work in the clinic to improve reproducibility and clinical trust.

Meng: I mean, taking those five different deep learning models—VNet, ResUNet, ReconNet, SAM-Medthree dee—and finding that VNet was the most robust one based on its geometric accuracy and radiomics stability metrics is a solid engineering result.

Tom: That’s what I like to see. It’s not just about which model has the highest test accuracy number; it's about which model stays consistent when you try to repeat the measurement, and VNet delivered that.

Jane: And they did this by testing it against expert contours and then having physicians look at those masks and give qualitative feedback across seven different domains to see if they felt it was useful in a real workflow.

Lu: That physician input is crucial because it shows where the AI might actually be failing in terms of clinical meaning, not just mathematical precision. It suggests that VNet's smoother boundaries were preferred by radiologists for peritumoral representation.

Meng: So, the main implication for us at the startup side is that we need to focus on that stability and clinical alignment over chasing the absolute highest accuracy score because consistency is what matters when you’re actually putting something in a hospital setting.

Tom: Exactly. The authors conclude that while VNet was the best model, it’s not a magic solution; it needs those outcome-linked metrics and physician input to be truly reliable for clinical use.

Jane: It sounds like the path forward isn't just about training better models, but about building an integrated system where the AI acts as a supportive tool for initial refinement rather than a final answer generator.

Lu: That’s the direction they point toward—developing platforms that integrate that VNet segmentation with things like click-point tracking so it becomes a more comprehensive diagnostic aid.

Meng: And I think that's where the real work is going to be, connecting the model output directly into a usable clinical software platform.

Tom: So, we’ve seen how they built this from data analysis all the way to clinical testing and concluded that VNet is a strong contender, but the next hurdle is making sure it actually fits smoothly into the doctor's day.

Mohammad R. Salmanpour, Sonya Falahati, Amir Hossein Pouria, Amin Mousavi, Somayeh Sadat Mehrnia, Morteza Alizadeh, Arman Gorji, Zeinab Farsangi, Alireza Safarian, Mehdi Maghsudi

Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada · Department of Radiology, University of British Columbia, Vancouver, BC, Canada · Technological Virtual Collaboration (TECVICO Corp.) · Amirkabir University of Technology · Abhar Branch Islamic Azad University Department of Computer Engineering Abhar · Department of Integrative Oncology Breast Cancer Research Center Motamed Cancer Institute ACECR Tehran · Department of Mathematics University of Isfahan Isfahan Iran · NAIRG Department of Neuroscience Hamadan University Medical Sciences Hamadan Iran · Rajaie Cardiovascular Medical and Research Center Iran University of Medical Science Tehran Iran · Department of Nuclear Medicine University Hospital Paracelsus Medical University Salzburg Austria · BC Cancer Vancouver Center Vancouver BC Canada

cs.CV

Submitted: 2025-10-19

Updated: 2026-11-12

Comments: 13 pages, 2 figures, and 2 tables

Code: https://github.com/MohammadRSalmanpour/Clinician-in-the-Loop-AI-Segmentation-for-CT-BasedPrognosis

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 77/100

The gist: Lung cancer is the leading cause of cancer death, with CT central to screening, prognosis, and treatment This study shows that VNet, integrated with SSL, can deliver accurate, reproducible…

Key concepts

Knowledge-to-Action (KTA) Framework
A structured approach used to guide the development of a clinician-in-the-loop pipeline. It ensures that research findings are translated into actionable steps for clinicians, focusing on improving reproducibility, prediction accuracy, and building clinical trust in AI tools.
VNet (Vision Network)
A deep learning model tested in this study that performed the best among five models. It achieved superior geometric accuracy and strong radiomics stability when segmenting lung cancer tumors from CT images, making it the most robust model for this application.
Semi-Supervised Learning (SSL)
A training framework used to improve model performance by combining labeled and unlabeled data. In this study, SSL consistently outperformed supervised learning because it provided more stable and reliable predictive performance during independent testing.

Terminology

Summary

Lung cancer is the leading cause of cancer death, with CT central to screening, prognosis, and treatment This study shows that VNet, integrated with SSL, can deliver accurate, reproducible, and clinically trusted tools for CTbased lung cancer prognosis

Methodological Framework

The research employed a Knowledge-to-Action (KTA) framework to guide the development of a clinician-in-the-loop pipeline to improve reproducibility, prediction, and clinical trust The study analyzed multi-center lung cancer CT data from 999 patients across 12 public datasets Five deep learning models (3D Attention U-Net, ResUNet, VNet, ReconNet, SAM-Med3D) were benchmarked against expert contours on whole and click-point cropped images Segmentation reproducibility was tested using 497 PySERA-extracted radiomics features with various statistical tests Six physicians qualitatively evaluated masks across 21 questions spanning seven domains to assess clinical meaningfulness and workflow integration

Model Performance and Stability

VNet-based segmentations achieved the best geometric accuracy (Dice 0.83±0.07, IoU 0.71±0.09) and radiomics stability (averaged correlation coefficient=0. p. 0

Where the paper says it

Results: VNet-based segmentations achieved the best geometric accuracy (Dice 0. 09), radiomics stability (averaged correlation coefficient=0.

76, ICC=0.65) Semi-supervised learning (SSL) consistently outperformed supervised learning (SL) across models The SSL configuration involving LASSO–LGBM–VNet achieved the highest test accuracy (0. p. 7
Where the paper says it

SSL models consistently achieved higher accuracy, recall, and F1-scores while also sustaining stronger AUC and specificity, particularly in configurations involving LASSO–LGBM– VNet and MIGR–ET–ground truth. Importantly, VNet stood out as the most powerful backbone among the segmentation models, delivering superior performance across both validation and test phases and showing greater robustness in SSL than in SL.

88 ± 0.003 accuracy, 0. p. 6
Where the paper says it

For instance, SSL with MI–MLP–ResUNet reached 0. 003 accuracy, 0.

83 ± 0.004 F1)

Clinical and Qualitative Insights

Qualitative assessments revealed that radiologists favored VNet for peritumoral representation and smoother boundaries Experts p referred AI to provide initial masks for refinement rather than replace physicians Physicians highlighted that AI masks often introduced noisy, unstable boundaries instead of biologically meaningful differences, limiting prognostic stability Overall assessments indicated no significant differences among the five models in diagnostic or prognostic success based on Friedman tests

Conclusion and Future Directions

The study concluded that VNet emerged as the most robust model, achieving superior geometric accuracy and strong radiomics reproducibility The findings demonstrate that SSL pipelines with VNet can enhance prognostic prediction, while clinical reliability requires outcome-linked metrics and physician input Future work includes developing a dedicated clinical software platform integrating VNet-based segmentation with click-point and gaze-tracking functionality The study addressed 10 of 11 barriers to routine adoption of AI in the clinic through its integrated evaluation framework

ACKNOWLEDGEMENTS

This study was supported by the Virtual Collaboration Group (VirCollab, www.vircollab.com) and the Technological Virtual Collaboration (TECVICO CORP. p. 9

Where the paper says it

ACKNOWLEDGEMENTS This study was supported by the Virtual Collaboration Group (VirCollab, www. com) and the Technological Virtual Collaboration (TECVICO CORP.

) based in Vancouver, Canada We gratefully acknowledge funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) Award RGPIN-2023-0357, and Discovery Horizons Grant DH-2025-00119, as well as the UBC Department of Radiology 2023 AI Fund

CONFLICT OF INTEREST

Drs. Mohammad R. p. 0

Where the paper says it

1 Click, Predict, Trust: Clinician-in-the-Loop AI Segmentation for Lung Cancer CT-Based Prognosis within the Knowledge-to-Action Framework Mohammad R. Salmanpour1,2,3,*, Sonya Falahati3 , Amir Hossein Pouria4 , Amin Mousavi5 , Somayeh Sadat Mehrnia6 , Morteza Alizadeh7 , Arman Gorji3,8, Zeinab Farsangi9 , Alireza Safarian9,10, Mehdi Maghsudi3 , Carlos Uribe1,2,11, Arman Rahmim1,2,# , Ren Yuan2,11,# 1Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada 2Department of Radiology, University of British Columbia, Vancouver, BC, Canada 3Technological Virtual Collaboration (TECVICO Corp.

Salmanpour, Mehdi Maghsudi, Arman Gorji, and Ms. p. 9
Where the paper says it

Mohammad R. Salmanpour, Mehdi Maghsudi, Arman Gorji, and Ms.

Sonya Falahati are affiliated with TECVICO Corp The other co-authors declare no relevant conflicts of interest or disclosures

DATA AND CODE AVAILABILITY

All code is publicly shared at: https://github.com/MohammadRSalmanpour/Clinician-in-the-Loop-AI-Segmentation-for-CT-BasedPrognosis/tree/main

REFERENCE

  1. Liao, X., et al., Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence‑assisted decision‑making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research. p. 9

    Where the paper says it

    com/MohammadRSalmanpour/Clinician-in-the-Loop-AI-Segmentation-for-CT-BasedPrognosis/tree/main REFERENCE 1. , Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence-assisted decision-making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research.

    BMJ Open, 2024. p. 9
    Where the paper says it

    , Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence-assisted decision-making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research. BMJ Open, 2024.

    14(9): p. e084398

  2. A. Fathi Jouzdani, et al., Crossing the AI Innovation‑to‑Application Chasm through Implementation Science: Using Concept Mapping in the Knowledge‑to‑Action Process to Identify and Overcome Barriers to Routine Use of Effective AI in Medical Imaging. p. 0

    Where the paper says it

    INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1]. The Knowledge-to-Action (KTA) framework within implementation science [2] provides such a roadmap, emphasizing that innovations must extend beyond methodological performance to overcome systemic, clinical, and workflow challenges.

    ArXiv, 2025

  3. Hassan, M., A. Kushniruk, and E. Borycki, Barriers to and Facilitators of Artificial Intelligence Adoption in Health Care: Scoping Review. p. 9

    Where the paper says it

    com/MohammadRSalmanpour/Clinician-in-the-Loop-AI-Segmentation-for-CT-BasedPrognosis/tree/main REFERENCE 1. , Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence-assisted decision-making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research.

    JMIR Hum Factors, 2024. p. 8
    Where the paper says it

    5 ResUNet 2. DISCUSSION This study shows that while DL segmentation models vary in geometric accuracy, radiomics stability, and predictive performance, their clinical utility becomes comparable once a certain level of reliability and consistency is achieved across these factors.

    11: p. e48633

  4. Ahmed, M.I., et al., A Systematic Review of the Barriers to the Implementation of Artificial Intelligence in Healthcare. p. 0

    Where the paper says it

    Keywords: Lung Cancer CT Prognosis; Deep Learning Segmentation; Radiomics Analysis; Semi-Supervised Learning; Clinician-inthe-Loop AI; Knowledge-to-Action Framework 1. INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1].

    Cureus, 2023. p. 9
    Where the paper says it

    ) based in Vancouver, Canada. We gratefully acknowledge funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) Award RGPIN-2023-0357, and Discovery Horizons Grant DH-2025-00119, as well as the UBC Department of Radiology 2023 AI Fund.

    15(10): p. e46454

  5. Ramadan, O.M.E., et al., Facilitators and barriers to AI adoption in nursing practice: a qualitative study of registered nurses' perspectives. BMC Nursing, 2024. p. 9

    Where the paper says it

    , Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence-assisted decision-making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research. BMJ Open, 2024.

    23(1): p. 891

  6. Petersson, L., et al., Challenges to implementing artificial intelligence in healthcare: a qualitative interview study with healthcare leaders in Sweden. p. 9

    Where the paper says it

    com/MohammadRSalmanpour/Clinician-in-the-Loop-AI-Segmentation-for-CT-BasedPrognosis/tree/main REFERENCE 1. , Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence-assisted decision-making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research.

    BMC Health Services Research, 2022. p. 0
    Where the paper says it

    1 Click, Predict, Trust: Clinician-in-the-Loop AI Segmentation for Lung Cancer CT-Based Prognosis within the Knowledge-to-Action Framework Mohammad R. Salmanpour1,2,3,*, Sonya Falahati3 , Amir Hossein Pouria4 , Amin Mousavi5 , Somayeh Sadat Mehrnia6 , Morteza Alizadeh7 , Arman Gorji3,8, Zeinab Farsangi9 , Alireza Safarian9,10, Mehdi Maghsudi3 , Carlos Uribe1,2,11, Arman Rahmim1,2,# , Ren Yuan2,11,# 1Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada 2Department of Radiology, University of British Columbia, Vancouver, BC, Canada 3Technological Virtual Collaboration (TECVICO Corp.

    22(1): p. 850

  7. Nair, M., et al., A comprehensive overview of barriers and strategies for AI implementation in healthcare: Mixed‑method design. p. 0

    Where the paper says it

    The Knowledge-to-Action (KTA) framework within implementation science [2] provides such a roadmap, emphasizing that innovations must extend beyond methodological performance to overcome systemic, clinical, and workflow challenges. Previous studies[3-8]have highlighted the significant barriers to AI adoption in healthcare, including infrastructure issues, clinical skepticism, and the need for human-centered decision-making in AI applications.

    PLOS ONE, 2024. p. 9
    Where the paper says it

    , Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence-assisted decision-making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research. BMJ Open, 2024.

    19(8): p. e0305949

  8. Adler-Milstein, J., et al., Meeting the Moment: Addressing Barriers and Facilitating Clinical Adoption of Artificial Intelligence in Medical Diagnosis. p. 0

    Where the paper says it

    Keywords: Lung Cancer CT Prognosis; Deep Learning Segmentation; Radiomics Analysis; Semi-Supervised Learning; Clinician-inthe-Loop AI; Knowledge-to-Action Framework 1. INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1].

    NAM Perspect, 2022

  9. Thandra, K.C., et al., Epidemiology of lung cancer. p. 0

    Where the paper says it

    1 Click, Predict, Trust: Clinician-in-the-Loop AI Segmentation for Lung Cancer CT-Based Prognosis within the Knowledge-to-Action Framework Mohammad R. Salmanpour1,2,3,*, Sonya Falahati3 , Amir Hossein Pouria4 , Amin Mousavi5 , Somayeh Sadat Mehrnia6 , Morteza Alizadeh7 , Arman Gorji3,8, Zeinab Farsangi9 , Alireza Safarian9,10, Mehdi Maghsudi3 , Carlos Uribe1,2,11, Arman Rahmim1,2,# , Ren Yuan2,11,# 1Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada 2Department of Radiology, University of British Columbia, Vancouver, BC, Canada 3Technological Virtual Collaboration (TECVICO Corp.

    Contemp Oncol (Pozn), 2021. 25(1): p. 45-52

  10. Corn, B.W. and D.B. Feldman, Cancer statistics, 2025: A hinge moment for optimism to morph into hope? p. 3

    Where the paper says it

    The extracted radiomic features are categorized into three major groups: 50 first-order intensity features, which capture basic statistics such as mean, variance, skewness, and kurtosis (including subsets of local intensity, intensity-based statistics, intensity histogram, and intensityvolume histogram); 29 shape and morphological features, which quantify geometric characteristicssuch as volume, surface area, compactness, and sphericity; and 408 texture features, which describe spatial intensity patterns and intra-tumoral heterogeneity. The texture group is further divided into 150 Gray-Level Co-occurrence Matrix (GLCM) features, 96 GrayLevel Run Length Matrix (GLRLM) features, 48 Gray-Level Size Zone Matrix (GLSZM) features, 48 Gray-Level Distance Zone Matrix (GLDZM) features, 15 Neighborhood Gray-Tone Difference Matrix (NGTDM) features, 51 Neighboring Gray-Level Dependence Matrix (NGLDM) features, and 10 moment-invariant features.

    CA: A Cancer Journal for Clinicians, 2025. p. 1
    Where the paper says it

    2 Lung cancer remains the leading cause of cancer-related mortality worldwide (>1. 8 million deaths annually) [9], with ~124,000 deaths projected in the United States by 2025 [10].

    75(1): p. 7-9

  11. Gałecki, S., et al., Assessing the Prognosis of Patients with Metastatic or Recurrent Non‑small Cell Lung Cancer in the Era of Immunotherapy and Targeted Therapy. p. 0

    Where the paper says it

    Guided by the Knowledge-to-Action framework, this study develops a clinician-in-the-loop pipeline to improve reproducibility, prediction, and clinical trust. Methods: Multi-center lung cancer CT data from 999 patients across 12 public datasets were analyzed.

    in The Latest Developments and Challenges in Biomedical Engineering. p. 0
    Where the paper says it

    Salmanpour1,2,3,*, Sonya Falahati3 , Amir Hossein Pouria4 , Amin Mousavi5 , Somayeh Sadat Mehrnia6 , Morteza Alizadeh7 , Arman Gorji3,8, Zeinab Farsangi9 , Alireza Safarian9,10, Mehdi Maghsudi3 , Carlos Uribe1,2,11, Arman Rahmim1,2,# , Ren Yuan2,11,# 1Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada 2Department of Radiology, University of British Columbia, Vancouver, BC, Canada 3Technological Virtual Collaboration (TECVICO Corp. ), Vancouver, BC, Canada 4Amirkabir University of Technology, Department of Computer Engineering, Tehran, Iran 5Department of Computer, Abhar Branch, Islamic Azad University, Abhar, Iran 6Department of Integrative Oncology, Breast Cancer Research Center, Motamed Cancer Institute, ACECR, Tehran, Iran 7Department of Mathematics, University of Isfahan, Isfahan, Iran 8NAIRG, Department of Neuroscience, Hamadan University of Medical Sciences, Hamadan, Iran 9Rajaie Cardiovascular Medical and Research Center, Iran University of Medical Science, Tehran, Iran 10Department of Nuclear Medicine, University Hospital, Paracelsus Medical University, Salzburg, Austria 11BC Cancer, Vancouver Center, Vancouver, BC, Canada *Corresponding Author: msalman@bccrc.

    2024

  12. Kelly, C.J., et al., Key challenges for delivering clinical impact with artificial intelligence. p. 0

    Where the paper says it

    INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1]. The Knowledge-to-Action (KTA) framework within implementation science [2] provides such a roadmap, emphasizing that innovations must extend beyond methodological performance to overcome systemic, clinical, and workflow challenges.

    BMC Medicine, 2019. p. 0
    Where the paper says it

    Salmanpour1,2,3,*, Sonya Falahati3 , Amir Hossein Pouria4 , Amin Mousavi5 , Somayeh Sadat Mehrnia6 , Morteza Alizadeh7 , Arman Gorji3,8, Zeinab Farsangi9 , Alireza Safarian9,10, Mehdi Maghsudi3 , Carlos Uribe1,2,11, Arman Rahmim1,2,# , Ren Yuan2,11,# 1Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada 2Department of Radiology, University of British Columbia, Vancouver, BC, Canada 3Technological Virtual Collaboration (TECVICO Corp. ), Vancouver, BC, Canada 4Amirkabir University of Technology, Department of Computer Engineering, Tehran, Iran 5Department of Computer, Abhar Branch, Islamic Azad University, Abhar, Iran 6Department of Integrative Oncology, Breast Cancer Research Center, Motamed Cancer Institute, ACECR, Tehran, Iran 7Department of Mathematics, University of Isfahan, Isfahan, Iran 8NAIRG, Department of Neuroscience, Hamadan University of Medical Sciences, Hamadan, Iran 9Rajaie Cardiovascular Medical and Research Center, Iran University of Medical Science, Tehran, Iran 10Department of Nuclear Medicine, University Hospital, Paracelsus Medical University, Salzburg, Austria 11BC Cancer, Vancouver Center, Vancouver, BC, Canada *Corresponding Author: msalman@bccrc.

    17(1): p. 195

  13. Aberle, D.R., et al., The National Lung Screening Trial: overview and study design. p. 1

    Where the paper says it

    Large randomized trials—the U. National Lung Screening Trial (NLST) and the Dutch–Belgian NELSON trial—established low-dose CT as the only screening modality proven to reduce lung cancer mortality [13, 14].

    Radiology, 2011. p. 0
    Where the paper says it

    1 Click, Predict, Trust: Clinician-in-the-Loop AI Segmentation for Lung Cancer CT-Based Prognosis within the Knowledge-to-Action Framework Mohammad R. Salmanpour1,2,3,*, Sonya Falahati3 , Amir Hossein Pouria4 , Amin Mousavi5 , Somayeh Sadat Mehrnia6 , Morteza Alizadeh7 , Arman Gorji3,8, Zeinab Farsangi9 , Alireza Safarian9,10, Mehdi Maghsudi3 , Carlos Uribe1,2,11, Arman Rahmim1,2,# , Ren Yuan2,11,# 1Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada 2Department of Radiology, University of British Columbia, Vancouver, BC, Canada 3Technological Virtual Collaboration (TECVICO Corp.

    258(1): p. 243-53

  14. de Koning, H.J., et al., Reduced Lung‑Cancer Mortality with Volume CT Screening in a Randomized Trial. p. 1

    Where the paper says it

    Large randomized trials—the U. National Lung Screening Trial (NLST) and the Dutch–Belgian NELSON trial—established low-dose CT as the only screening modality proven to reduce lung cancer mortality [13, 14].

    N Engl J Med, 2020

  15. Facktor, M.A., et al., Artificial intelligence–aided lung cancer screening in routine clinical practice: A pilot of LungFlag at Geisinger. p. 0

    Where the paper says it

    Keywords: Lung Cancer CT Prognosis; Deep Learning Segmentation; Radiomics Analysis; Semi-Supervised Learning; Clinician-inthe-Loop AI; Knowledge-to-Action Framework 1. INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1].

    Journal of Clinical Oncology, 2024. p. 0
    Where the paper says it

    Salmanpour1,2,3,*, Sonya Falahati3 , Amir Hossein Pouria4 , Amin Mousavi5 , Somayeh Sadat Mehrnia6 , Morteza Alizadeh7 , Arman Gorji3,8, Zeinab Farsangi9 , Alireza Safarian9,10, Mehdi Maghsudi3 , Carlos Uribe1,2,11, Arman Rahmim1,2,# , Ren Yuan2,11,# 1Department of Basic and Translational Research, BC Cancer Research Institute, Vancouver, BC, Canada 2Department of Radiology, University of British Columbia, Vancouver, BC, Canada 3Technological Virtual Collaboration (TECVICO Corp. ), Vancouver, BC, Canada 4Amirkabir University of Technology, Department of Computer Engineering, Tehran, Iran 5Department of Computer, Abhar Branch, Islamic Azad University, Abhar, Iran 6Department of Integrative Oncology, Breast Cancer Research Center, Motamed Cancer Institute, ACECR, Tehran, Iran 7Department of Mathematics, University of Isfahan, Isfahan, Iran 8NAIRG, Department of Neuroscience, Hamadan University of Medical Sciences, Hamadan, Iran 9Rajaie Cardiovascular Medical and Research Center, Iran University of Medical Science, Tehran, Iran 10Department of Nuclear Medicine, University Hospital, Paracelsus Medical University, Salzburg, Austria 11BC Cancer, Vancouver Center, Vancouver, BC, Canada *Corresponding Author: msalman@bccrc.

    42(16 suppl): p. e13604-e13604

  16. Farsad, M., FDG PET/CT in the Staging of Lung Cancer. p. 1

    Where the paper says it

    Because CT is high-resolution, widely available, and relatively low cost, it is central to screening and follow-up, addressing infrastructure (i), workflow (vii), and clinical skepticism (ii) while mitigating financial concerns (v) [15]. Although PET/CT remains essential for staging [16], CT is the natural platform for embedding AI-enabled radiomics and prognostic pipelines [17, 18].

    Curr Radiopharm, 2020. 13(3): p. 195-203

  17. Zhang, Y.-P., et al., Artificial intelligence-driven radiomics study in cancer: the role of feature engineering and modeling. p. 0

    Where the paper says it

    Keywords: Lung Cancer CT Prognosis; Deep Learning Segmentation; Radiomics Analysis; Semi-Supervised Learning; Clinician-inthe-Loop AI; Knowledge-to-Action Framework 1. INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1].

    Military Medical Research, 2023. p. 9
    Where the paper says it

    ) based in Vancouver, Canada. We gratefully acknowledge funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) Award RGPIN-2023-0357, and Discovery Horizons Grant DH-2025-00119, as well as the UBC Department of Radiology 2023 AI Fund.

    10(1): p. 22

  18. Bagci, U., X. Chen, and J.K. Udupa, Hierarchical scale-based multiobject recognition of 3‑D anatomical structures. p. 8

    Where the paper says it

    These discrepancies seldom corresponded to biologically relevant regions but rather reflected random noise, which likely contributed to the inferior prognostic performance of AI masks compared to ground truth. While AI sometimes captured broader peritumoral regions or additional structures not included in ground truth masks, this was inconsistent and often contaminated with nontumoral tissue, reducing reliability.

    IEEE Trans Med Imaging, 2012. p. 0
    Where the paper says it

    Keywords: Lung Cancer CT Prognosis; Deep Learning Segmentation; Radiomics Analysis; Semi-Supervised Learning; Clinician-inthe-Loop AI; Knowledge-to-Action Framework 1. INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1].

    31(3): p. 777-89

  19. Clèrigues, A., et al., Acute ischemic stroke lesion core segmentation in CT perfusion images using fully convolutional neural networks. p. 2

    Where the paper says it

    ReconNet provides a lightweight and efficient alternative, while SAM-Med3D introduces recent advances in foundation models with strong generalization capabilities [67]. To ensure robustness, these networks were applied to both whole CT images and cropped (click-point–guided) tumor-centered volumes, enabling a direct evaluation of segmentation variability

    Comput Biol Med, 2019. 115: p. 103487

  20. Guzene, L., et al., Assessing Interobserver Variability in the Delineation of Structures in Radiation Oncology: A Systematic Review. Int J Radiat Oncol Biol Phys, 2023. p. 9

    Where the paper says it

    ) based in Vancouver, Canada. We gratefully acknowledge funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) Award RGPIN-2023-0357, and Discovery Horizons Grant DH-2025-00119, as well as the UBC Department of Radiology 2023 AI Fund.

    115(5): p. 1047-1060

  21. Mehrnia, S.S., et al., Landscape of 2D Deep Learning Segmentation Networks Applied to CT Scan from Lung Cancer Patients: A Systematic Review. p. 0

    Where the paper says it

    ca #Co-Last-Authors ABSTRACT Objective: Lung cancer is the leading cause of cancer death, with CT central to screening, prognosis, and treatment. Manual segmentation is variable and time-consuming, while deep learning offers automation but faces adoption barriers.

    J Imaging Inform Med, 2025

  22. Yao, I.Z., M. Dong, and W.Y.K. Hwang, Deep Learning Applications in Clinical Cancer Detection: A Review of Implementation Challenges and Solutions. p. 0

    Where the paper says it

    Keywords: Lung Cancer CT Prognosis; Deep Learning Segmentation; Radiomics Analysis; Semi-Supervised Learning; Clinician-inthe-Loop AI; Knowledge-to-Action Framework 1. INTRODUCTION For artificial intelligence (AI) in medical imaging to achieve clinical adoption, technical advances must be embedded within frameworks that address barriers to implementation [1].

    Mayo Clinic Proceedings: Digital Health, 2025

  23. Traverso, A., et al., Machine learning helps identifying volume‑confounding effects in radiomics. p. 3

    Where the paper says it

    Model selection was based on validation performance, and generalizability was further assessed through external testing on two independent datasets, such as LungCT-Diagnosis and NSCLC-Radiomics-Genomics (~10%). Segmentation performance was measured using the Dice similarity coefficient, IoU, and Hausdorff distance, all computed via the AllMetrics library, which standardizes machine learning evaluation [70].

    Phys Med, 2020. 71: p. 24-30

  24. Larue, R.T., et al., Quantitative radiomics studies for tissue characterization: a review of technology and methodological procedures. p. 0

    Where the paper says it

    The Knowledge-to-Action (KTA) framework within implementation science [2] provides such a roadmap, emphasizing that innovations must extend beyond methodological performance to overcome systemic, clinical, and workflow challenges. Previous studies[3-8]have highlighted the significant barriers to AI adoption in healthcare, including infrastructure issues, clinical skepticism, and the need for human-centered decision-making in AI applications.

    Br J Radiol, 2017. 90(1070): p. 20160665

  25. Alves, N., et al., Prediction Variability to Identify Reduced AI Performance in Cancer Diagnosis at MRI and CT. p. 4

    Where the paper says it

    Such physician-in-the-loop interpretation is not only about validating DL outputs but also about revealing how much margin or peritumoral region clinicians may need to consistently include for reliable diagnosis and prediction. In alignment with the KTA framework, this classification analysis addressed barriers of adaptability (iii), variability and reliability (xi), limitations in atypical or complex cases (ix), and workforce resistance (viii) by demonstrating that DL models can complement rather than r eplace clinical expertise.

    Radiology, 2023. p. 9
    Where the paper says it

    ) based in Vancouver, Canada. We gratefully acknowledge funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) Award RGPIN-2023-0357, and Discovery Horizons Grant DH-2025-00119, as well as the UBC Department of Radiology 2023 AI Fund.

    308(3): p. e230275

  26. Hershman, M., et al., Impact of Interobserver Variability in Manual Segmentation of Non‑Small Cell Lung Cancer (NSCLC) Applying Low-Rank Radiomic Representation on Computed Tomography. Cancers (Basel), 2021. p. 3

    Where the paper says it

    If discrepancies are found, further qualitative review and downstream classification or survival modeling will be performed to investigate and explain any variations. This framework strengthens the study's robustness and provides a generalizable model for assessing segmentation reliability across cancers and imaging modalities, advancing the integration of AI into clinical imaging practice.

    13(23).

  27. Maier-Hein, L., et al., Metrics reloaded: recommendations for image analysis validation. p. 5

    Where the paper says it

    1 Lesion Segmentation Results As shown in Figure 1, V-Net on cropped (click-point) images consistently outperforms the other models, all in terms of Dice score, Hausdorff Distance, and IoU, particularly in both the test and validation datasets. In the test dataset, VNet achieves the highest values for both metrics (Dice: 0.

    Nat Methods, 2024. p. 9
    Where the paper says it

    , Barriers and facilitators to implementing imaging-based diagnostic artificial intelligence-assisted decision-making software in hospitals in China: a qualitative study using the updated Consolidated Framework for Implementation Research. BMJ Open, 2024.

    21(2): p. 195-212

308(3): p. e230275

--- Page 6 ---

Pruning redundant reasoning tokens

The study shows that while DL segmentation models vary in geometric accuracy, radiomics stability, and predictive performance, their clinical utility becomes comparable once a certain level of reliability and consistency is achieved across these factors Among the five well-known models, VNet achieved the best testing Dice (0.83) and strong radiomics reproducibility ReconNet produced radiomics features with notable prognostic value despite weaker geometric scores Friedman tests showed no significant differences for q10 (diagnosis success), q11 (prediction success), or q12 (confidence)

How the training signal is built

The study utilized two complementary frameworks: supervised learning (SL) and semi-supervised learning (SSL) In the SSL framework, a logistic regression (LR) model trained on four folds generated pseudo-labels for unlabeled datasets Dimensionality reduction strategies, including feature selection algorithms and automated embedding algorithms, were applied to address high dimensionality A broad panel of 24 classifiers was used to evaluate each reduced feature set

Where the gains come from

The superior performance of SSL models was observed in independent test evaluations, where they showed significantly more stable and reliable predictive performance compared to SL The novel contribution lies in its integrated evaluation framework, which unites geometric metrics, radiomics reproducibility, survival modeling, and physician-centered assessments under the KTA paradigm VNet emerged as the most suitable model for clinical translation due to its balance of accuracy and clinical judgment alignment The qualitative analysis confirmed that clinicians prefer AI not as a replacement, but as a system that provides an initial mask for refinement or serves as a supportive tool VNet addresses barriers of infrastructure and integration (i), adaptability (iii), skepticism (ii), trust (iv), reliability (xi), and workflow disruption (vii)

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Clinical Trust Enhancement through Clinician-in-the-Loop Integration: The system can be designed to provide initial masks for refinement rather than replacing physicians, as experts p referred AI to provide initial masks for refinement rather than replace physicians.

  2. Improved Segmentation Accuracy via VNet Backbone: Utilizing the VNet architecture, which consistently outperforms the other models in terms of Dice score and IoU, can yield superior geometric accuracy (Dice 0.83±0.07).

  3. Enhanced Prognostic Reliability through Radiomics Stability: The system will incorporate a radiomics stability analysis to ensure features remain robust, as this testing target[s] trust (iv) and clinical distrust (x).

  4. Robust Prediction via Semi-Supervised Learning: Employing the SSL framework, specifically configurations like LASSO–LGBM–VNet, can achieve higher accuracy and AUC by leveraging unlabeled data that shows SSL consistently outperformed SL across models.

  5. Workflow Efficiency through Tumor-Centered Cropping: Implementing a tumor-centered region-ofinterest (ROI) cropping strategy preserved the largest connected component with a 5-voxel 3D margin, thereby retaining peritumoral context while minimizing background, which addresses workflow disruption (vii).

  6. Increased Interpretability via Confidence Maps: The system should generate confidence maps to improve transparency, as physicians noted that interpretabil[ity] tools, such as confidence or uncertainty maps, would improve transparency and trust.

Abstract

Lung cancer remains the leading cause of cancer mortality, with CT imaging central to screening, prognosis, and treatment. Manual segmentation is variable and time-intensive, while deep learning (DL) offers automation but faces barriers to clinical adoption. Guided by the Knowledge-to-Action framework, this study develops a clinician-in-the-loop DL pipeline to enhance reproducibility, prognostic accuracy, and clinical trust. Multi-center CT data from 999 patients across 12 public datasets were analyzed using five DL models (3D Attention U-Net, ResUNet, VNet, ReconNet, SAM-Med3D), benchmarked against expert contours on whole and click-point cropped images. Segmentation reproducibility was assessed using 497 PySERA-extracted radiomic features via Spearman correlation, ICC, Wilcoxon tests, and MANOVA, while prognostic modeling compared supervised (SL) and semi-supervised learning (SSL) across 38 dimensionality reduction strategies and 24 classifiers. Six physicians qualitatively evaluated masks across seven domains, including clinical meaningfulness, boundary quality, prognostic value, trust, and workflow integration. VNet achieved the best performance (Dice = 0.83, IoU = 0.71), radiomic stability (mean correlation = 0.76, ICC = 0.65), and predictive accuracy under SSL (accuracy = 0.88, F1 = 0.83). SSL consistently outperformed SL across models. Radiologists favored VNet for peritumoral representation and smoother boundaries, preferring AI-generated initial masks for refinement rather than replacement. These results demonstrate that integrating VNet with SSL yields accurate, reproducible, and clinically trusted CT-based lung cancer prognosis, highlighting a feasible path toward physician-centered AI translation.

Sources

Related papers