Investigation of a Novel Deep Learning-Based Computed Tomography Perfusion Mapping Framework for Functional Lung...

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Functional lung avoidance radiation therapy requires accurate regional lung function information.
• Current methods for acquiring pulmonary functional images are resource-intensive and technically challenging.

Purpose of the Study:

• To investigate the feasibility of synthesizing lung functional images using deep learning from CT scans.
• To develop and evaluate a deep learning framework for estimating regional lung function.

Main Methods:

• A deep learning framework, including image preparation, processing (with CT contrast enhancement), and a convolutional neural network (CNN), was used.
• The framework synthesized perfusion images from 3D CT scans of 42 patients.
• Ablation experiments assessed the impact of different framework components on performance.

Main Results:

• CT contrast enhancement in image processing was critical, with its removal causing the largest performance drop (~12%).
• CNN components (residual module, ROI attention, skip attention) were equally important, each removal causing a 3-5% decline.
• The proposed CNN improved performance by ~4% and computational efficiency by ~350% compared to U-Net.

Conclusions:

• Deep convolutional neural networks combined with image processing can effectively synthesize pulmonary perfusion from CT images.
• Image processing, particularly CT contrast enhancement, is vital for accurate perfusion synthesis.
• The developed framework offers a more accessible approach for functional lung avoidance radiation therapy planning.