An anthropomorphic diagnosis system of pulmonary nodules using weak annotation-based deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate lung nodule categorization in CT scans is crucial for early lung cancer detection and diagnosis.
• Current nodule classification methods require extensive high-quality annotations, limiting clinical applicability.
• Differentiating nodule grade and texture aids clinical decision-making in nodule management.

Purpose of the Study:

• To develop a deep learning (DL) based system for pulmonary nodule (PN) diagnosis using weak annotation data.
• To achieve performance comparable to full-annotation systems while reducing annotation burden.
• To enable efficient localization and differential diagnosis of PNs, including various subtypes.

Main Methods:

• Development of a DL system for PN classification (benign vs. malignant) trained on weak annotations.
• Integration of handcrafted shape features using the ball-scale transform technique.
• 5-fold cross-validation on the LIDC-IDRI and an in-house dataset.

Main Results:

• Achieved an Area Under Curve (AUC) of 0.938 for PN localization and 0.912 for differential diagnosis on the LIDC-IDRI dataset (814 cases).
• Achieved an AUC of 0.943 for PN localization and 0.815 for differential diagnosis on the in-house dataset (822 cases).
• Demonstrated capability to differentiate PNs with diverse labels like ground-glass, part-solid, and solid nodules.

Conclusions:

• The developed DL system efficiently localizes and diagnoses pulmonary nodules using weak annotations.
• The system shows comparable performance to full-annotation systems, addressing limitations of current methods.
• The system's efficiency in resource-limited environments suggests potential for clinical translation in lung cancer diagnosis.