Deep learning for segmentation, detection, and grading of lumbar disc herniation and related structures on MRI: a systematic review

Background Deep learning (DL) has been investigated for segmentation, localization, detection, and grading on lumbar magnetic resonance imaging (MRI). These tasks support different aspects of lumbar disc herniation assessment, and technical performance does not establish clinical benefit. Methods PubMed, Embase, the Cochrane Library, Scopus, Web of Science, CINAHL, and IEEE Xplore were searched, with the final search conducted on 11 June 2026. Eligible applications included segmentation, localization, detection, classification, and grading; segmentation was not mandatory. Two reviewers independently screened and extracted data. Findings were synthesized narratively and the modified QUADAS-2 assessment was interpreted cautiously. Results The original inventory contained 17 reports. Reported Dice similarity coefficients for disc-related segmentation ranged from approximately 0.82 to 0.962, and intersection over union values reached 0.976 in selected tasks. These study-specific results were not pooled. Studies were heterogeneous and predominantly retrospective, with limited prospective clinical validation. Conclusion DL shows potential for lumbar MRI segmentation and related detection and grading tasks. Current evidence does not establish readiness for routine clinical deployment. Reproducible searches, task-specific appraisal, independent external validation, and prospective evaluation of clinical outcomes are needed.