An Anthropomorphic Diagnosis System of Pulmonary Nodules using Weak Annotation-Based Deep Learning

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Pulmonary nodules (PNs) require accurate diagnosis for effective lung cancer management.
• Traditional diagnostic systems rely on extensive manual annotations, which are time-consuming and costly.
• Developing efficient and accurate PN diagnosis systems is crucial for clinical practice.

Approach:

• A deep learning (DL) system was developed for anthropomorphic diagnosis of pulmonary nodules (PNs) using weak annotation data.
• The system employs DL models for classifying PNs (benign vs. malignant) and integrates handcrafted shape features via ball-scale transform.
• Multi-scale networks and fusion techniques enhance classification performance across diverse nodule types.

Key Points:

• Achieved high performance in PN localization (AUC=0.938) and differential diagnosis (AUC=0.912) on the LIDC-IDRI dataset.
• Demonstrated strong results on an in-house dataset (AUC=0.943 for localization, AUC=0.815 for diagnosis).
• The system integrates expert knowledge (handcrafted features) into DL models, improving robustness and efficiency.

Conclusions:

• The developed DL system achieves performance comparable to full-annotation systems for PN diagnosis.
• Weak annotations significantly reduce annotation time and cost, making the system suitable for resource-limited settings.
• The system shows potential for clinical translation due to its robustness and efficiency in diagnosing pulmonary nodules.