Semi-supervised training using cooperative labeling of weakly annotated data for nodule detection in chest CT

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Detection

Background:

• Accurate annotations are crucial for training machine learning algorithms.
• Medical image annotation is time-consuming and expensive, limiting training dataset size.
• Clinically produced data are often weakly annotated, lacking machine-readable details.

Purpose of the Study:

• To propose a cooperative labeling method for training machine learning algorithms using weakly annotated medical imaging data.
• To incorporate a wider range of clinical data by leveraging existing weakly annotated scans.
• To increase the size of training datasets for medical machine learning applications.

Main Methods:

• A multi-stage pseudo-labeling approach was developed.
• An initial network trained on expert annotations generates pseudo-labels for weakly annotated data.
• Cross-checking annotations and combining datasets create a larger, higher-fidelity training set for a computer-aided detection system.

Main Results:

• The cooperative labeling method was evaluated for nodule detection in chest CT scans.
• Inclusion of weakly labeled data improved the competitive performance metric (CPM) by 5% when expert annotations were limited.
• The approach effectively merges weakly and fully annotated data to enhance model training.

Conclusions:

• The proposed method effectively merges weakly and fully annotated datasets for algorithm training.
• This approach addresses the challenge of limited annotated data in medical imaging.
• Leveraging weakly labeled clinical data can significantly enlarge training datasets for improved diagnostic AI.