Related Experiment Video
Updated: Jun 29, 2025
10:26
Author Spotlight: A 3D Digital Model for the Diagnosis and Treatment of Pulmonary Nodules
Published on: May 19, 2023
1.8K
Enhancing a deep learning model for pulmonary nodule malignancy risk estimation in chest CT with uncertainty
Dré Peeters1, Natália Alves2, Kiran V Venkadesh2
1Diagnostic Imaging Analysis Group, Medical Imaging Department, Radboud University Medical Center, Geert Grooteplein Zuid 10, 6525 GA, Nijmegen, the Netherlands. dre.peeters@radboudumc.nl.
European Radiology
|March 27, 2024
Summary
Uncertainty estimation significantly improves the safety of deep learning (DL) algorithms for pulmonary nodule malignancy risk assessment. This method identifies cases where the DL algorithm performs poorly, enhancing clinical implementation and trustworthiness.
Area of Science:
- Medical Imaging Analysis
- Artificial Intelligence in Radiology
- Machine Learning for Healthcare
Background:
- Deep learning (DL) algorithms are increasingly used for pulmonary nodule malignancy risk estimation.
- A key limitation of current DL algorithms is their inability to reliably quantify prediction uncertainty.
- This lack of uncertainty estimation poses challenges for safe clinical adoption.
Purpose of the Study:
- To evaluate the impact of integrating an uncertainty estimation method into a DL algorithm for pulmonary nodule malignancy risk.
- To determine if uncertainty estimation can identify cases where the DL algorithm's performance is compromised.
Main Methods:
- A DL algorithm for nodule malignancy risk was modified to include uncertainty estimation.
- Uncertainty thresholds were derived from the Danish Lung Cancer Screening Trial (DLCST) dataset.
- External validation was performed on a separate clinical dataset, assessing DL performance on certain and uncertain nodule groups.
Main Results:
- The DL algorithm demonstrated significantly lower performance (AUC) in the uncertain nodule group compared to the certain group on both datasets.
- Uncertain cases were characterized by larger benign nodules and a higher proportion of part-solid and non-solid nodules.
- The uncertainty estimation method effectively identified these challenging cases.
Conclusions:
- Integrated uncertainty estimation successfully identified pulmonary nodules where the DL algorithm's performance was significantly diminished.
- This approach is crucial for enhancing the safety and reliability of DL algorithms in clinical practice.
- Uncertainty quantification is pivotal for the trustworthy implementation of AI in medical diagnostics.
