Deep neural network for water/fat separation: Supervised training, unsupervised training, and no training

Area of Science:

• Medical Imaging
• Computational Science

Background:

• Water/fat separation is a critical step in various imaging techniques.
• Current algorithms like IDEAL require specific initialization, which can be a limitation.
• Deep Neural Networks (DNNs) have emerged as a potential solution for automated water/fat separation.

Purpose of the Study:

• To investigate the application of DNNs for optimizing water/fat separation.
• To compare the efficacy of supervised and unsupervised DNN training methods for this task.

Main Methods:

• Proposed two novel DNN-based water/fat separation methods: unsupervised training of DNN (UTD) and DNN with no training.
• UTD utilized the physical forward problem as its cost function.
• The no-training DNN method employed physical cost and backpropagation for direct reconstruction.

Main Results:

• All DNN approaches, including supervised, unsupervised, and no-training methods, produced water/fat separation results consistent with the IDEAL algorithm.
• The agreement was particularly strong when IDEAL was used with appropriate initialization.

Conclusions:

• Unsupervised deep neural networks provide a viable and effective approach to solving the water/fat separation problem.
• DNN methods offer a promising alternative to traditional algorithms, potentially reducing reliance on manual initialization.