Diff-FFT: Synthetic Lumbar Magnetic Images From Frequency Domain-Based

Magnetic resonance imaging is playing an increasingly important role in the diagnosis of lumbar disc disease. It can provide high-resolution soft tissue contrast, helping doctors accurately identify disc lesions, inflammation and changes in other related structures. However, the use of MRI is limited due to its high cost and significant operating and processing times. More importantly, MRI examinations are not allowed for people with pacemakers, other metal objects in the body such as cochlear implants and metal teeth, or people in critical condition or claustrophobic, as it may cause injury. For this reason, this paper proposes a frequency domain-based method for synthesizing lumbar magnetic images. First, we introduce a multi-scale frequency channel attention module to model global features, which helps the model more effectively process wide-area information and improve its accuracy and detail performance in generating lumbar magnetic resonance images. Then, a dual-resolution frequency domain-based FFN module is proposed. This module selectively retains important frequency information through a learnable quantization matrix W. The inverse method of JPEG compression. Experimental results show that this method is accurate and can construct MR images that are very close to the reference MR images, and it also outperforms the other three state-of-the-art methods.