Multitask Deep Learning Reconstruction and Localization of Lesions in Limited Angle Diffuse Optical Tomography

Area of Science:

• Biomedical optics
• Medical imaging
• Computational imaging

Background:

• Diffuse optical tomography (DOT) uses near-infrared light for tissue assessment but faces ill-posed reconstruction challenges.
• Limited-angle DOT increases complexity and often results in artifacts, hindering accurate lesion detection and localization.
• Conventional reconstruction methods are computationally intensive, limiting real-time applications.

Purpose of the Study:

• To develop a fast and accurate deep learning-based image reconstruction method for DOT.
• To improve both image reconstruction quality and lesion localization accuracy using multitask learning.
• To address the scarcity of real-world training data by utilizing physics-based simulations and transfer learning.

Main Methods:

• A novel multitask deep learning framework incorporating spatial-wise attention and a distance transform-based loss function.
• Physics-based simulations to generate synthetic datasets for training.
• Transfer learning to bridge the domain gap between simulated and real-world sensor data.

Main Results:

• The proposed method achieves faithful reconstruction and localization of lesions in real-time.
• Demonstrated ability to reconstruct multiple cancer lesions accurately.
• Multitask learning significantly improves reconstruction sharpness and accuracy compared to single-task methods.

Conclusions:

• Deep learning, particularly multitask learning with attention and specialized loss functions, offers a promising solution for fast and accurate DOT image reconstruction.
• The integration of simulation and transfer learning effectively overcomes data scarcity issues.
• This approach enables real-time, high-fidelity imaging for improved lesion detection and characterization.