Automated neural foraminal stenosis grading via task-aware structural representation learning

Neural foraminal stenosis (NFS) is the most common spinal disease in elderly patients, greatly affecting their quality of life. Efficient and accurate grading of NFS is extremely vital for physicians as it offers patients a timely and targeted treatment according to different grading levels. However, current clinical practice relies on physicians' visual inspection and manual grading of neural foramina (NF), which brings the annoying inefficiency and inconsistency. A fully automated system is highly desirable but faces many technical challenges (e.g., the inefficiency in localizing NF candidates, and the severe ambiguities in grading). In this paper, an automated and accurate localization and grading clinical framework is proposed. By our framework, both localization and grading tasks are handled as multi-class classification problem: two-class classification (NF/non-NF) and four-class classification (normal/slight/marked/severe). To achieved it, a newly proposed saliency-biased Ncuts (SBNcuts) is utilized for efficient localization, and a novel task-aware structural representation learning (TASRL) model is developed for accurate localization and grading. Specifically, SBNcuts creatively incorporates saliency map as a preliminary guess of NF's locations to refine the generated possible NF candidates with the preserved intact structure of NF. TASRL incorporates task labels (e.g., NF object label and four NFS grade labels) into manifold learning to obtain a discriminative, low-dimensional, and structural image representation, which enables similar appearance sharing among images with the same task label and different appearance among images with different task labels. The superior performance in localization and grading, with very high ( > 0.89) accuracy, specificity, sensitivity, and F-measure, have been demonstrated by experiments on 110 subjects. With our method, physicians could offer an efficient and consistent clinical grading for NFS.