Implementation of a nested U-Net architecture within the nnU-Net framework for spine segmentation in MRI

Spinal metastases in oncologic patients are a widely diffused pathology and can lead to vertebral fractures, severely threatening the patients’ quality of life. In this scope, the METASTRA EU Research Project aims to develop a clinical decision support tool in order to evaluate the fracture risks of the spine and provide concise and personalised indications and risk evaluations to the medical staff. This work has been developed within METASTRA and focuses on the implementation of a pipeline for the automated segmentation of Intervertebral Discs from MR images via deep network architectures. To achieve this goal, a nested U-Net architecture, the U²-Net, was employed within the nnU-Net framework, the current state-of-the-art framework for medical image segmentation, that automatically selects the appropriate hyperparameters for each case-study. The model was trained and evaluated on both T1-Weighted and T2-Weighted spinal MRI publicly available in the SPIDER dataset, aiming to identify the optimal modality for the Intervertebral Disc segmentation and to evaluate the U²-Net performances. The analysis of the results showed T2-Weighted images as better suited for discs segmentation and specifically the 3D U²-Net configuration evaluated on T2-Weighted MR images yielded a mean Dice similarity coefficient of 0.89. The performance of the model has been compared with a baseline derived from the U-Net already available in the nnU-Net framework, trained on the same dataset, that yielded a Dice similarity coefficient score of 0.85, lower than U²-Net. The model was also trained on the T2-weighted images to segment the vertebrae aiming to enable the registration of the Intervertebral Disc segmentation masks obtained from T2-Weighted MR images with the vertebrae segmentation masks from CT images, that notably provide better delineation of bone tissues. In this case, the optimal configuration was found to be the 2D U²-Net, yielding a Dice similarity coefficient of 0.93.