DMCA-GAN: Dual Multilevel Constrained Attention GAN for MRI-Based Hippocampus Segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Accurate hippocampus segmentation is crucial for studying brain activity and neurological disorders.
• Challenges include the hippocampus's small size and low contrast in MRI scans, hindering precise segmentation.

Purpose of the Study:

• To develop an effective deep learning model for precise hippocampus segmentation in MRI.
• To address limitations of existing methods in handling noise and feature learning for small, low-contrast structures.

Main Methods:

• Proposed a dual multilevel constrained attention Generative Adversarial Network (DMCA-GAN) for MRI hippocampus segmentation.
• Incorporated a dual-GAN backbone for spatial information compensation and a multilayer information constraint unit for noise suppression.
• Implemented a multiscale feature attention mechanism to refine boundary segmentation and improve robustness.

Main Results:

• The dual-GAN backbone improved Dice coefficient (DSC) by 5.95% over the baseline.
• The multilayer information constraint unit enhanced feature sensitivity by 5.39% and reduced network overfitting.
• DMCA-GAN achieved a DSC of 90.53% on the Medical Segmentation Decathlon (MSD) dataset, outperforming the backbone by 3.78%.

Conclusions:

• The proposed DMCA-GAN effectively balances noise suppression and feature learning for accurate hippocampus segmentation.
• This method demonstrates superior performance and robustness compared to existing approaches, offering a valuable tool for neurological research.