Research map / Reinforcement learning / Policy gradient and actor-critic
Deep Reinforcement Learning for Early Diagnosis of Lung Cancer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formulates lung nodule follow-up vs. biopsy decisions as a belief-MDP optimal stopping problem and solves it with value-iteration deep RL (EarlyStop-RL) on NLST, outperforming Lung-RADS and Brock models.
Problem
Deciding after each LDCT screening whether to continue follow-up or order a biopsy/diagnose lung cancer, balancing delayed diagnosis against false positives and negatives.
Approach
Models lung cancer natural history as a POMDP, converts it to a belief MDP with HMM belief updates, and uses Snell-envelope optimal stopping with a model-free deep Q-learning (value iteration) solver; stopping regions are convex, and a linear SVM approximates the threshold boundaries for interpretable deployment.
Key results
- EarlyStop-RL achieves a 60.88% early diagnosis rate versus 35.47% for Lung-RADS and 33.64% for the Brock model.
- EarlyStop-RL reaches FNR 1.33%, F1 0.85 and MCC 0.80, versus Lung-RADS FNR 12.00%, F1 0.66, MCC 0.52.
- Overall NRI is 0.24 versus Lung-RADS (p=0.00) and 0.10 versus Brock model (p=0.03).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EarlyStop-RL instead of Lung-RADS | Lung cancer screening diagnosis · NLST test set | false-positive rate: 29.47% -> 12.85% | Helped |
| Use EarlyStop-RL instead of Lung-RADS | Lung cancer screening diagnosis · NLST test set | false-negative rate: 12.00% -> 1.33% | Helped |
| Use EarlyStop-RL instead of Lung-RADS | Lung cancer screening diagnosis · NLST test set | early diagnosis rate: 35.47% -> 60.88% | Helped |
| Use EarlyStop-RL instead of Lung-RADS | Lung cancer screening diagnosis · NLST test set | F1 score: 0.66 -> 0.85 | Helped |
| Use EarlyStop-RL instead of Lung-RADS | Lung cancer screening diagnosis · NLST test set | MCC: 0.52 -> 0.80 | Helped |
| Use EarlyStop-RL instead of Brock Model | Lung cancer screening diagnosis · NLST test set | false-positive rate: 13.10% -> 12.85% | Helped |
| Use EarlyStop-RL instead of Brock Model | Lung cancer screening diagnosis · NLST test set | false-negative rate: 4.67% -> 1.33% | Helped |
| Use EarlyStop-RL instead of Brock Model | Lung cancer screening diagnosis · NLST test set | early diagnosis rate: 33.64% -> 60.88% | Helped |
| Use EarlyStop-RL instead of Brock Model | Lung cancer screening diagnosis · NLST test set | F1 score: 0.83 -> 0.85 | Helped |
| Use EarlyStop-RL instead of Brock Model | Lung cancer screening diagnosis · NLST test set | MCC: 0.77 -> 0.80 | Helped |
| Use EarlyStop-RL instead of Google AI Model | Lung cancer screening diagnosis · NLST test set | false-positive rate: 9.00% -> 12.85% | Hurt |
| Use EarlyStop-RL instead of Google AI Model | Lung cancer screening diagnosis · NLST test set | false-negative rate: 4.40% -> 1.33% | Helped |
Surprising findings
- Under the NLST trial protocol, only 17.73% of test patients received an early diagnosis, implying over 80% face delayed diagnosis risk.
Limitations
- Observation model relies only on clinical risk factors such as nodule diameter rather than learned image features.
- No prospective study or external validation was conducted.
- NLST data access requires permission, limiting reproducibility.
Experimental setup
Datasets: National Lung Screening Trial (NLST)
Benchmarks: NLST
Metrics: false-positive rate false-negative rate F1 score Matthews correlation coefficient early diagnosis rate net reclassification index
Baselines: Lung-RADS Brock model Google AI model
Builds on: Snell envelope optimal stopping POMDP Hidden Markov Model filter Q-learning
Closest papers by meaning
- Deep Reinforcement Learning for Cost-Effective Medical Diagnosis (ICLR 2023 · similarity 0.86)
- Cross-Modal Interactive Perception Network with Mamba for Lung Tumor Segmentation in PET-CT Images (CVPR 2025 · similarity 0.82)
- LungNoduleAgent: A Collaborative Multi-Agent System for Precision Diagnosis of Lung Nodules (AAAI 2026 · similarity 0.82)
- Patho-R1: A Multimodal Reinforcement Learning-Based Pathology Expert Reasoner (AAAI 2026 · similarity 0.82)
- Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective (ICML 2025 · similarity 0.82)
- CMR-RD: Long-Tailed Adaptive VLM for Explainable CMR Diagnosis (CVPR 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.