Weakly-supervised Learning Based Spine Instance Segmentation for MRI Planning

Abstract

Magnetic Resonance Imaging (MRI) spine planning involves setting several positioning lines, termed localizer, through the intervertebral discs (IVDs) of interest to enable axial scans. Deep learning models that generate IVD masks facilitate the automation of MRI spine planning workflow. However, training an automatic IVD segmentation model typically requires extensive data across multiple spine regions and labor-intensive annotation of segmentation masks. To reduce the annotation burden, we propose a novel weakly supervised learning based two-stage training strategy for IVD instance segmentation models. During the first supervised learning stage, the model is trained on a public dataset with segmentation annotations. During the second weakly supervised stage, the pre-trained model is fine-tuned on a local spine dataset annotated only with localizers. To achieve the weakly supervised learning, a distance map generated by localizer is leveraged to refine the predicted mask’s principal axis, enabling the model to adapt to the data distribution of the local dataset. We conducted experiments by combining our strategy with several popular models on lumbar, cervical, and thoracic spine datasets. To quantify the performance of localizer predictions, we designed evaluation metrics based on angular deviation, centroid accuracy, and Hausdorff distance. Results demonstrate that the successful integration of the proposed strategy with various baseline instance segmentation models, resulting in notable improvements in performance across images with different signal-to-noise ratios and anatomical regions. In conclusion, the proposed weakly supervised method enhances generalizability across diverse scenarios and holds promise for improving the clinical adoption of automatic spine planning in MRI examinations.