Adaptive machine learning method for photoacoustic computed tomography based on sparse array sensor data.

Area of Science:

• Biomedical Imaging
• Machine Learning
• Medical Technology

Background:

• Photoacoustic computed tomography (PACT) is a rapidly advancing non-invasive imaging technique with potential in early disease diagnosis.
• High-element-density detector arrays are crucial for quality PACT images, but are often limited by cost, manufacturing, and system constraints.
• Sparse detector arrays in PACT can lead to artifacts and reduced image quality.

Purpose of the Study:

• To develop an adaptive machine learning method for predicting and complementing photoacoustic sensor channel data from sparse array sampling.
• To improve the quality of reconstructed PACT images by addressing data sparsity and artifacts.
• To offer a cost-effective and user-friendly solution for PACT imaging.

Main Methods:

• An adaptive machine learning model combining XGBoost and a neural network (SS-net) was developed.
• A tunable parameter was employed to balance XGBoost and SS-net outputs, enhancing generalization across different dataset sizes.
• The method predicts and complements sparse photoacoustic sensor data before image reconstruction.

Main Results:

• The proposed method demonstrated superior performance in simulations, phantom, and in vivo experiments.
• Compared to existing methods, significant improvements were observed in Structural Similarity Index Measure (SSIM) and R-squared values.
• Specifically, SSIM increased by up to 21.46% and median R² by up to 84.1% with in vivo data.

Conclusions:

• The developed model effectively predicts missing photoacoustic sensor data for sparse ring-shaped arrays in PACT.
• The method significantly suppresses artifacts and enhances image quality compared to linear interpolation and deep learning alternatives.
• This approach requires no large pre-trained image datasets, using sensor data directly, and has broad potential for clinical PACT applications.