Image-to-Image Translation for Simplified MRI Muscle Segmentation

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) show promise in radiology but require extensive labeled data for training.
• The need for large labeled datasets hinders the practical application of DNNs in medical imaging.
• Automated segmentation of pathological muscle tissue in T1-weighted MR images is challenging.

Purpose of the Study:

• To develop a method for simplifying medical image segmentation without requiring labeled data.
• To create an "easier-to-segment" intermediate image representation using unpaired image-to-image translation.
• To evaluate the effectiveness of this approach for segmenting pathological muscle tissue in thigh MR images.

Main Methods:

• Utilized an unpaired image-to-image translation approach.
• Developed a novel domain-specific loss formulation.
• Applied the method to segment pathological muscle tissue in T1-weighted MR images.
• Investigated fully automated segmentation approaches.

Main Results:

• The proposed method significantly improved performance for supervised segmentation techniques.
• Achieved comparable results using a basic unsupervised segmentation approach.
• Demonstrated the creation of an "easier-to-segment" intermediate image representation.

Conclusions:

• Unpaired image-to-image translation with a novel loss function can effectively simplify medical image segmentation tasks.
• This technique reduces the dependency on labeled data, enabling practical applications of DNNs in radiology.
• The approach shows potential for both supervised and unsupervised segmentation of pathological tissues.