Exploring synthetic datasets for computer-aided detection: a case study using phantom scan data for enhanced lung...

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Artificial intelligence in diagnostics

Background:

• Clinical datasets for training AI are often limited, costly, and may contain biases.
• Synthetic datasets offer a privacy-preserving, cost-effective alternative.
• Computer-aided detection (CADe) systems require robust training data.

Purpose of the Study:

• To present a method for training machine learning algorithms using synthetic datasets for CADe systems.
• To evaluate the effectiveness of using physical phantom data and unlabeled clinical data for training.
• To improve the performance of lung nodule CADe systems.

Main Methods:

• Utilized computed tomography (CT) scans of an anthropomorphic phantom with manufactured lesions.
• Augmented phantom data with randomized and parameterized variations.
• Incorporated unlabeled clinical data to mitigate domain differences.
• Applied the trained algorithm to the false positive reduction stage of a lung nodule CADe system.

Main Results:

• Achieved 90% sensitivity at eight false positives per scan for lung nodule detection.
• Demonstrated a 6% performance increase by augmenting a clinical training set with phantom data.
• Validated the effectiveness of synthetic data and unlabeled clinical data in training.

Conclusions:

• Synthetic datasets are scalable and can significantly improve CADe performance.
• The proposed method effectively utilizes synthetic data for training machine learning algorithms.
• This approach is particularly beneficial when labeled clinical data is scarce or biased.