A Dataset of Pulmonary Lesions With Multiple-Level Attributes and Fine Contours

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Computational Pathology

Background:

• Lung cancer diagnosis is critical but hindered by limited datasets.
• Developing robust computer-aided diagnosis (CAD) systems requires comprehensive data.

Purpose of the Study:

• Introduce a novel dataset of pulmonary lesions with detailed annotations.
• Facilitate the development and evaluation of CAD systems for lung cancer.

Main Methods:

• Dataset creation with fine contour and nine attribute annotations.
• Statistical analysis (chi-square test) of attribute correlations.
• Development and testing of multi-attribute classification models in 2D, 2.5D, and 3D.
• Integration and verification of attention mechanisms to enhance model performance.

Main Results:

• Established the structure and detailed annotations of the pulmonary lesion dataset.
• Demonstrated the utility of the dataset for multi-attribute classification tasks.
• Showcased performance improvements using attention mechanisms in CAD models.
• Validated correlations between lesion attributes and pathological findings.

Conclusions:

• The new dataset effectively addresses data scarcity in lung cancer research.
• The proposed CAD models show promise for improved lung cancer diagnosis.
• Attention mechanisms offer a viable strategy for enhancing CAD system performance.