Automatic Grading of Nucleus Pulposus Degeneration for spinal endoscopic surgery based on Ultrasound Signal analysis by an enhanced ViT

Lumbar Disc Herniation is caused by degeneration of tissues such as nucleus pulposus (NP), and can lead to the compression of nerve roots. The real-time grading of the NP degeneration during spinal endoscopic surgery can help surgeons locate the lesion region, so that degenerated tissues can be removed precisely and completely. The objective of this study is to explore an automated degeneration grading method based on the analysis of ultrasound signals acquired from an endoscopic transducer using an enhanced Vision Transformer (ViT). We proposed a Patch Tokenizer (PT) and a Channel-Independent (CI) training method to enhance the performance of ViT. The PT splits the input signals into several patches, and then a fully connection layer was applied to transform the patches into embeddings. The CI training method can augment the training data, and alleviate the occasionality of the prediction in the testing stage, thus enhancing the overall model performance. A custom-developed forward-oriented ultrasound endoscopic transducer was applied to collect signals in the experiment. 152 NP samples were collected from 24 patients with degenerative disease of the lumbar spine, and were labeled into 3 grades of degeneration. The proposed enhanced ViT model with PT and CI could achieve an average accuracy of 87.33% with a sensitivity of 81.49% and a specificity of 90.68% on the testset. The accuracy and specificity of 3 grades were all above 85%. The ablation study also showed the effectiveness of PT and CI. It was demonstrated that the proposed model would provide a high confident reference to the surgeons, and had the potential to serve in real-time automatic tissue identification during spinal surgeries.