Automated Segmentation Method for Low Field 3D Stomach MRI Using Transferred Learning Image Enhancement Network

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate segmentation of abdominal organs, particularly the stomach, is challenging for MRI-guided radiotherapy due to low CT contrast and limitations of low-field MRI.
• Existing 3D segmentation models trained on low-field MRI yield poor performance for radiotherapy planning.
• Using high-field MRI data for training introduces domain shift issues, hindering network learning.

Purpose of the Study:

• To propose a 3D low-field MRI stomach segmentation method using transfer learning and image enhancement.
• To overcome the domain shift problem between high and low-field MRI images.
• To improve the segmentation accuracy of deep neural networks for clinical radiotherapy applications.

Main Methods:

• Employed Cycle Generative Adversarial Network (CycleGAN) to learn mapping between high and low-field MRI, addressing domain shift.
• Utilized CycleGAN-generated high-field MRI data as extended training datasets for low-field MRI.
• Trained a 3D Res-Unet segmentation network incorporating residual modules to mitigate gradient disappearance.

Main Results:

• Achieved a 2.5% improvement in Dice coefficient compared to baseline methods.
• Reduced over-segmentation by 0.7% and under-segmentation by 5.5%.
• Improved segmentation sensitivity by 6.4%.

Conclusions:

• The proposed CycleGAN-based transfer learning method effectively enhances low-field MRI stomach segmentation accuracy.
• This approach overcomes domain shift challenges, enabling better integration of MRI data into radiotherapy planning.
• The improved segmentation performance is critical for accurate stomach localization, tracking, and treatment planning in radiotherapy.