|
In clinical practice, nerve root compromise mainly refers to diseases in which the protrusion of the nucleus pulposus of the spinal cord leads to irritation or compression of the spinal cord and spinal nerve roots, causing corresponding neurological symptoms. Magnetic Resonance Imaging (MRI) of the lumbar spine is an important tool to grade the nerve root compromise. The traditional diagnosis method mainly relies on manual identification by doctors, so it is inefficient. At present, the computer aided diagnosis system based on deep learning has been widely used in the medical field, and the diagnosis accuracy of many diseases has surpassed that of clinicians. Vision Transformers have been successfully applied to several computer vision problems, achieving state-of-the-art results. These advancements in computer vision have resulted in a growing interest in Transformers in the medical image processing field. Vision Transformers can capture global context, but lack the local receptive field of CNNs and is relatively weak in modeling local spatial context. Many researchers have tried and achieved preliminary results, but there is still a long way to go before it can be widely used in the medical field like CNN. This paper aims to expand the application of visual Transformer in the field of medical image processing, using two convolution methods, atrous convolution embedding and depthwise atrous convolution projection, to improve the Swin Transformer for automatic classification of nerve root compromise in MR images of Lumbar Spine. The introduction of convolution effectively improves the performance of the Swin Transformer model, and enables the network to better take into account global and local context information. Our method achieved 0.8234 accuracy and 0.8185 F1 score on test set, which improved by 1.31% and 1.94% respectively compared to the baseline model.
|
