Prior-aware autoencoders for lung pathology segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of lung pathologies in Computed Tomography (CT) images is crucial for disease screening.
• Challenges in segmentation arise from diverse pathology characteristics and visual similarity to surrounding tissues.
• Existing methods struggle with reliable automatic lesion delineation.

Purpose of the Study:

• To develop a deep learning framework for improved lung pathology segmentation in CT images.
• To leverage a Normal Appearance Autoencoder (NAA) model to generate prior information for segmentation.
• To enhance the accuracy of segmenting pulmonary nodules, Non-Small Cell Lung Cancer (NSCLC), and COVID-19 lesions.

Main Methods:

• A deep learning framework incorporating a Normal Appearance Autoencoder (NAA) was proposed.
• The NAA model learns healthy lung tissue distribution to reconstruct pathology-free images.
• Pathology-free reconstructions provided prior information (shape, location) integrated into a segmentation network.

Main Results:

• The proposed framework demonstrated superior performance compared to baseline segmentation models across three pathology types.
• Significant improvements in Dice coefficient were observed: 0.038 for lung nodules, 0.101 for NSCLC, and 0.041 for COVID-19 lesions.
• The NAA model successfully generated reliable prior knowledge guiding more accurate delineations.

Conclusions:

• The Normal Appearance Autoencoder (NAA) model effectively generates reliable prior knowledge for lung pathologies.
• Integrating NAA-derived prior information into segmentation networks significantly enhances delineation accuracy.
• The proposed framework offers a promising approach for precise lung pathology segmentation in clinical settings.