Chronic pain originating from the lumbar spinal canal is a significant health concern and often challenging to diagnose due to the limitations of conventional imaging techniques. While MRI provides high-resolution anatomical detail, it fails to capture the functional abnormalities associated with chronic pain. PET/MRI, which combines metabolic activity from PET with anatomical imaging from MRI, has shown promise in detecting spinal lesions linked to pain. However, identifying these lesions remains labour-intensive due to their subtle and often indistinct nature. This study aimed to develop an automated approach for segmenting and detecting abnormalities in PET images of the lumbar spinal canal. First, the spinal canal was segmented from MRI scans using nnU-Net. After that, SUV values in PET data were extracted based on the segmented spinal canal. A combination of explainable image analysis techniques—thresholding, graph-based analysis, and random forest classifier—were employed to identify abnormal metabolic activity. The resulting probability maps highlight potential lesions, assisting radiologists and nuclear medicine specialists in diagnosis. The proposed model was validated against expert annotations and synthetic abnormalities, successfully detecting 51 abnormalities and missing one. This enhancement of the detection process has the potential to streamline clinical workflows and improve diagnostic accuracy for chronic pain patients.