Robust deep learning from incomplete annotation for accurate lung nodule detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Deep learning for pulmonary nodule detection in LDCT scans is crucial for lung cancer diagnosis.
• Current methods require extensive, meticulously annotated datasets, increasing costs and time.
• Incomplete annotations present challenges for training effective deep learning models.

Purpose of the Study:

• To develop an innovative algorithm, FULFIL, for pulmonary nodule detection using incompletely annotated datasets.
• To reduce annotation costs by having annotators label only confident nodules.
• To enable self-adaptive learning and annotation completion for robust deep learning models.

Main Methods:

• Utilized Graph Convolution Network (GCN) for self-adaptive annotation completion by discovering relationships between annotated and unannotated nodules.
• Employed a teacher-student framework for self-adaptive learning on the completed annotation dataset.
• Designed a Dual-Views loss function to enhance feature robustness and model generalization.

Main Results:

• Achieved a sensitivity of 0.574 at 0.125 False positives per scan (FPs/scan) using only 10% instance-level annotations on the LUNA dataset.
• Outperformed comparative methods by 7.00%.
• Demonstrated performance comparable to human experts in experimental comparisons.

Conclusions:

• FULFIL effectively leverages incomplete pulmonary nodule datasets to develop robust deep learning models.
• The algorithm significantly reduces annotation costs while maintaining high detection performance.
• FULFIL shows promise as a tool to assist in lung nodule detection and lung cancer diagnosis.