FDU-Net: Deep Learning-Based Three-Dimensional Diffuse Optical Image Reconstruction.

Area of Science:

• Biomedical optics
• Medical imaging
• Artificial intelligence in healthcare

Background:

• Near-infrared diffuse optical tomography (DOT) shows promise for breast cancer detection but faces clinical translation challenges.
• Conventional finite element method (FEM)-based DOT reconstruction is slow and struggles with accurate lesion contrast recovery.

Purpose of the Study:

• To develop and evaluate a deep learning-based model (FDU-Net) for fast and accurate 3D DOT image reconstruction in breast cancer imaging.
• To overcome the limitations of conventional FEM-based DOT reconstruction methods.

Main Methods:

• Developed FDU-Net, a deep learning model integrating fully connected, convolutional encoder-decoder, and U-Net subnets for end-to-end 3D DOT reconstruction.
• Trained FDU-Net on digital phantoms with diverse inclusions and evaluated its performance against FEM and other deep learning approaches in 400 simulated cases.
• Assessed model generalizability using multi-focal, irregularly shaped inclusions and validated with a real patient breast tumor measurement.

Main Results:

• FDU-Net significantly enhanced image quality and lesion contrast recovery compared to FEM-based methods and a prior deep learning network.
• The model achieved over a four-order-of-magnitude acceleration in computational time post-training.
• FDU-Net demonstrated generalizability to unseen inclusion shapes and successfully reconstructed a real patient's breast tumor.

Conclusions:

• The deep learning-based FDU-Net offers superior performance and computational efficiency for DOT breast cancer imaging compared to conventional methods.
• FDU-Net has the potential to enable real-time, accurate lesion characterization, assisting in clinical diagnosis and management of breast cancer.