| Autor: |
Irina Balzer, Malin Mühlemann, Moritz Jokeit, Ishaan Singh Rawal, Jess G. Snedeker, Mazda Farshad, Jonas Widmer |
| Jazyk: |
angličtina |
| Rok vydání: |
2022 |
| Předmět: |
|
| Zdroj: |
Computer Methods and Programs in Biomedicine Update, Vol 2, Iss , Pp 100081- (2022) |
| Druh dokumentu: |
article |
| ISSN: |
2666-9900 |
| DOI: |
10.1016/j.cmpbup.2022.100081 |
| Popis: |
Background: This work evaluates the feasibility, development, and validation of a machine learning pipeline that includes all tasks from MRI input to the segmentation and grading of the intervertebral discs in the lumbar spine, offering multiple different radiological gradings of degeneration as quantitative objective output. Methods: The pipelines’ performance was analysed on 1′000 T2-weighted sagittal MRI. Binary outputs were assessed with the harmonic mean of precision and recall (DSC) and the area under the precision-recall curve (AUC-PR). Multi-class output scores were averaged and complemented by the Top-2 categorical accuracy. The processing success rate was evaluated on 10′053 unlabelled MRI scans of lumbar spines. Results: The midsagittal plane selection achieved an DSC of 74,80% ± 2,99% and an AUC-PR score of 81.71% ± 2.72% (96.91% Top-2 categorical accuracy). The segmentation network obtained a DSC of 91.80% ± 0.44%. The Pfirrmann grading of intervertebral discs in the midsagittal plane was classified with a DSC of 64.08% ± 3.29% and an AUC-PR score of 68.25% ± 6.00% (91.65% Top-2 categorical accuracy). Disc herniations achieved a DSC of 61.57% ± 3.39% and an AUC-PR score of 66.86% ± 5.03%. The cranial endplate defects reached a DSC of 49.76% ± 3.45% and 52.36% ± 1.98% AUC-PR (slightly superior predictions of caudal endplate defect). The binary classifications for the caudal Schmorl's nodes obtained a DSC of 91.58% ± 2.25% with an AUC-PR metric of 96.69% ± 1.58% (similar performance for cranial Schmorl's nodes). Spondylolisthesis was classified with a DSC of 89.03% ± 2.42% and an AUC-PR score of 95.98% ± 1.82%. Annular Fissures were predicted with a DSC of 78.09% ± 7.21% and an AUC-PR score of 86.31% ± 7.45%. Intervertebral disc classifications in the parasagittal plane achieved an equivalent performance. The pipeline successfully processed 98.53% of the provided sagittal MRI scans. Conclusions: The present deep learning framework has the potential to aid the quantitative evaluation of spinal MRI for an array of clinically established grading systems. |
| Databáze: |
Directory of Open Access Journals |
| Externí odkaz: |
|
