Identification of potentially painful disc fissures in magnetic resonance images using machine-learning modelling.

Purpose: It is suggested that non-specific low back pain (LBP) can be related to nerve ingrowth along granulation tissue in disc fissures, extending into the outer layers of the annulus fibrosus. Present study aimed to investigate if machine-learning modelling of magnetic resonance imaging (MRI) data can classify such fissures as well as pain, provoked by discography, with plausible accuracy and precision.

Methods: The study was based on previously collected data from 30 LBP patients (age = 26-64 years, 11 males). Pressure-controlled discography was performed in 86 discs with pain-positive discograms, categorized as concordant pain-response at a pressure ≤ 50 psi and for each patient one negative control disc. The CT-discograms were used for categorization of fissures. MRI values and standard deviations were extracted from the midsagittal part and from 5 different sub-regions of the discs. Machine-learning algorithms were trained on the extracted MRI markers to classify discs with fissures extending into the outer annulus or not, as well as to classify discs as painful or non-painful.

Results: Discs with outer annular fissures were classified in MRI with very high precision (mean of 10 repeated testings: 99%) and accuracy (mean: 97%) using machine-learning modelling, but the pain model only demonstrated moderate diagnostic accuracy (mean accuracy: 69%; precision: 71%).

Conclusion: The present study showed that machine-learning modelling based on MRI can classify outer annular fissures with very high diagnostic accuracy and, hence, enable individualized diagnostics. However, the model only demonstrated moderate diagnostic accuracy regarding pain that could be assigned to either a non-sufficient model or the used pain reference.