AIBSD: Deep learning approach to address spatial systematic errors in diffusion tensor imaging

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Diffusion Tensor Imaging (DTI) metrics are vital for assessing tissue microstructure.
• Spatial systematic errors in DTI arise from inaccurate b-matrix estimation.
• Current correction methods necessitate time-consuming phantom scans, limiting clinical use.

Purpose of the Study:

• To develop and validate AIBSD, a deep learning method for correcting DTI b-matrices.
• To eliminate systematic errors in DTI without requiring additional phantom measurements.
• To improve the accuracy and clinical feasibility of DTI metrics.

Main Methods:

• A convolutional neural network (CNN) with a dual-encoder architecture was trained on 130 DTI datasets.
• The CNN predicted corrected b-matrices using reference data from the BSD-DTI method.
• DTI metrics were calculated using standard, BSD-DTI, and AIBSD (with phantom and in vivo data).

Main Results:

• AIBSD significantly reduced systematic errors, as shown by decreased FA values in phantom data.
• In vivo analysis demonstrated strong agreement between AIBSD and BSD-DTI (CCC > 0.993).
• AIBSD using in vivo data achieved accuracy comparable to phantom-based methods without phantom scans.

Conclusions:

• AIBSD effectively corrects spatial systematic errors in DTI using in vivo data.
• The method eliminates the need for additional phantom scans, enhancing clinical DTI feasibility.
• AIBSD improves DTI metric accuracy and streamlines workflows for research and clinical applications.