An Optimized Framework of QSM Mask Generation Using Deep Learning: QSMmask-Net.

Area of Science:

• Medical Imaging
• Neuroimaging
• Biophysics

Background:

• Quantitative susceptibility mapping (QSM) reconstructs tissue magnetic susceptibility distributions.
• Accurate mask generation is critical for QSM to minimize artifacts and errors.
• Current mask generation methods can introduce variability in susceptibility values.

Purpose of the Study:

• To introduce QSMmask-net, a deep neural network for precise QSM mask generation.
• To evaluate the performance of QSMmask-net against existing methods.
• To assess the impact of QSMmask-net on susceptibility value accuracy.

Main Methods:

• Development of QSMmask-net, a deep learning model for automated QSM mask creation.
• Comparison of QSMmask-net's Dice score with other mask generation techniques.
• Validation of QSMmask-net masks using simulations, healthy controls, and hemorrhagic lesions.

Main Results:

• QSMmask-net achieved the highest Dice score among tested mask generation methods.
• Susceptibility values derived from QSMmask-net masks showed minimal differences from manual masks in simulations and healthy controls.
• Strong linear correlation was observed between QSMmask-net and manual masks for hemorrhagic lesions.

Conclusions:

• Mask generation significantly impacts QSM susceptibility value estimation.
• QSMmask-net provides high-quality masks comparable to manual methods with reduced labor.
• The proposed method enhances QSM applicability by enabling efficient, expert-level mask creation.