Transformer- and joint learning-based dual-domain networks for undersampled MRI segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) is vital in clinical practice but faces challenges like long acquisition times and manual annotation needs.
• Current solutions address MRI reconstruction and segmentation separately, potentially limiting overall performance.
• Exploring end-to-end training for undersampled MRI reconstruction and segmentation is crucial to enhance efficiency and accuracy.

Purpose of the Study:

• Introduce a novel Transformer- and Joint Learning-based Dual-domain Network (TJLD-Net) for undersampled MRI segmentation.
• To enhance feature recognition and improve segmentation precision by integrating reconstruction and segmentation processes.
• To leverage attention mechanisms for better contextual information learning in MR images.

Main Methods:

• Developed a Dual-domain Network (TJLD-Net) integrating Transformer and Joint Learning approaches.
• Employed end-to-end training for simultaneous undersampled MRI reconstruction and segmentation.
• Incorporated an attention mechanism to enhance feature representation and contextual understanding.

Main Results:

• TJLD-Net demonstrated significantly higher segmentation performance compared to joint and serial baseline models across three datasets (CHAOS, ATLAS, SKM-TEA).
• Achieved substantial improvements in Dice scores, with notable gains on the CHAOS dataset (up to 9.87%) and SKM-TEA dataset (up to 14.83%).
• The model effectively overcomes performance degradation in automated segmentation for accelerated, low-quality MRI.

Conclusions:

• TJLD-Net offers a promising solution for accurate and reliable undersampled MRI segmentation.
• Jointly training reconstruction and segmentation in an end-to-end manner enhances performance.
• The proposed method addresses critical limitations in accelerated MRI segmentation, improving clinical applicability.