Introduction: The role of multifidus muscle dysfunction is increasingly recognized as a contributor to chronic low back pain (cLBP) and is an emerging therapeutic target. Currently, assessment of multifidus impairment relies on manual, subjective radiographic evaluation of fatty infiltration.
Methods: Using a dataset of 962 axial T2-weighted lumbar MRI scans from an open-source repository with expert segmentation, we designed a two-stage computer vision framework. The first stage employed a segmentation technique to isolate adipose and muscle tissues surrounding the spinous processes, focusing on the multifidus muscle. In the second stage, the images were processed through an adversarial Inception-v3 model that performed multiclass classification, assigning each case to one of four atrophy grades (0–3).
Results: Our two-stage pipeline demonstrated high accuracy in grading multifidus atrophy across a diverse set of axial T2-weighted MRI scans from multiple sources. The segmentation model achieved a mean average precision mAP of 0.61 and was reliably able to circumscribe relevant musculature. The classification model achieved an overall accuracy of 0.69, along with a macro-averaged F1 score of 0.73, a precision score of 0.68, and a recall score of 0.79. The distribution of atrophy severities was broad across the cohort, and when compared with a subset of expert-annotated scans, the system showed strong agreement while improving reproducibility significantly.
Conclusion : We introduce a fully automated, two-stage computer vision-based tool for the assessment of multifidus atrophy, offering a reliable measure of multifidus dysfunction while reducing interobserver variability. The algorithm’s integration into clinical workflows would enable earlier detection of muscle deterioration and provide timely intervention opportunities to halt the progression of cLBP.
.jpg)
