A Deep Transfer Learning Framework for Speed-of-Sound Aberration Correction in Full-Ring Photoacoustic Tomography.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Speed-of-sound (SoS) variations cause significant artifacts in full-ring photoacoustic tomography (PAT), limiting its clinical applications.
• Artifacts degrade image accuracy and hinder the reliable interpretation of PAT reconstructions.

Purpose of the Study:

• To develop and evaluate a deep neural framework for end-to-end artifact correction in PAT.
• To improve the accuracy and quality of full-view PAT reconstructions affected by SoS heterogeneities.

Main Methods:

• A transfer learning-based deep neural framework was designed, combining a ResNet-50 encoder with a deconvolutional decoder.
• A two-phase curriculum learning strategy was employed: pretraining on uniform SoS mismatches and fine-tuning on heterogeneous SoS fields.
• The framework was validated using numerical models, physical phantoms, and in vivo experiments.

Main Results:

• The proposed framework demonstrated substantial improvements over conventional back-projection and U-Net baselines.
• Key metrics such as mean squared error, structural similarity index measure, and Pearson correlation coefficient showed significant gains.
• The method achieved a fast average inference time of 17 ms per frame.

Conclusions:

• The deep learning framework effectively reduces the sensitivity of full-ring PAT to SoS inhomogeneity.
• The approach significantly enhances full-view reconstruction quality, offering a promising solution for clinical PAT applications.