Evaluation of an interpretable deep-learning model for the automated plan review of intensity-modulated radiation therapy
-
By
-
Yuhan Fan
-
Jiawen Shang
-
Ke Zhang
-
Zhihui Hu
-
Zhiqiang Liu
-
Hui Yan
-
Peng Huang
-
July 15, 2026
Clinical Scorecard: Assessment of a Transparent Deep Learning Framework for Automated Review of Intensity-Modulated Radiation Therapy Plans
At a Glance
| Category | Detail |
|---|
| Condition | Intensity-Modulated Radiation Therapy (IMRT) |
| Key Mechanisms | Automated anomaly detection using deep learning and interpretable feature ranking methods. |
| Target Population | Cancer patients receiving radiotherapy. |
| Care Setting | Radiotherapy clinics. |
Key Highlights
- Autoencoder (AE) model achieved an AUC of 0.98 for anomaly detection.
- Feature Perturbation Interpretation (FPI) method provides insights into feature impacts.
- AE outperformed classic detection models in identifying anomalous plans.
- Automated methods assist in quality assurance of IMRT plans.
- Interpretability of AI models is crucial for clinical trust.
Guideline-Based Recommendations
Diagnosis
- Radiotherapy plans should be reviewed for adherence to clinical guidelines.
Management
- Utilize automated detection algorithms to assist in plan review processes.
Monitoring & Follow-up
- Implement a physics plan review process that includes automated checks.
Risks
- High complexity of IMRT plans necessitates robust quality assurance to prevent errors.
Patient & Prescribing Data
Patients undergoing neo-adjuvant, definitive, adjuvant, or palliative radiotherapy.
Quality control processes are essential to ensure safe and effective treatment delivery.
Clinical Best Practices
- Conduct thorough manual verification of radiotherapy plans by experienced physicists.
- Incorporate machine learning methods to enhance the plan review process.
- Ensure interpretability of AI models to facilitate clinical acceptance.
Related Resources & Content
