Deep Learning-Based Generation of 4D Reference Lung Volumes for Personalized Respiratory Function Assessment.

Area of Science:

• Medical imaging
• Artificial intelligence
• Pulmonary medicine

Background:

• Lung function varies significantly within and between individuals.
• Accurate diagnosis and treatment of respiratory diseases require personalized reference points.
• Current methods may not fully capture dynamic lung behavior.

Purpose of the Study:

• To establish a personalized reference lung dynamic for individual patients.
• To predict a patient's hypothetical "healthy" respiratory dynamics.
• To enable comparison of actual lung function with a personalized healthy baseline.

Main Methods:

• Utilizing dynamic 3D magnetic resonance imaging (MRI) of the lungs.
• Employing an AI-based approach combining a convolutional autoencoder and an LSTM recurrent neural network.
• Generating dynamic lung models and predicting healthy respiratory motion.

Main Results:

• Successfully generated personalized "healthy" hypothetical respiratory dynamics.
• Developed a method to create static lung representations from real data.
• Captured temporal dependencies in respiratory deformations for motion generation.

Conclusions:

• AI-driven personalized lung dynamics can serve as a valuable baseline for respiratory health.
• This approach facilitates more precise and individualized respiratory disease assessment.
• Future applications include improved diagnosis, treatment, and follow-up for lung conditions.