Deep learning-based lower back pain classification and detection from T2-weighted magnetic resonance images

• Main Author: Al-Rubaye, M. (Mustafa)
• Format: Master Thesis
• Language: English
• Published: University of Oulu 2023
• Subjects: Lääketieteen tekniikka (tietotekniikka); Northern Finland
• Online Access: http://jultika.oulu.fi/Record/nbnfioulu-202302141142

Abstract. Lower back pain (LBP) is a common physiological condition that affects 50–80% of the adult population at some point in their lives. For example, the economic load of LBP in Sweden was estimated to be approx. at C740 million in 2011. In LBP diagnostics, magnetic resonance imaging (MRI) is often used. MRI is used to visualize the structures in the lumbar region of the spine such as disks, bones, and spaces between the vertebral bones where nerves pass through. The lumbar spine refers to the lowest five vertebrae and intervertebral discs of the spine. MRI provides a detailed picture of the lumbar spine to get visual confirmation of any abnormalities potentially related to LBP to support the diagnosis process. The goal of this thesis was to investigate visual patterns related to LBP in T2-weighted MR images measured with a fast spin-echo sequence on a GE Healthcare Signa HDxt 1.5 T MRI system. A convolutional neural network was used to classify MRIs into symptomatic and asymptomatic cases and to develop a fully automated pain prediction process. A total of 526 MRI examinations with supporting pain questionnaires from the Northern Finland Birth Cohort 1966 (NFBC1966) were used. Three different datasets were created for the experiments: i) a dataset with mid-sagittal slices from the center of the spine from each examination, ii) a dataset with mid-sagittal slices and its immediate neighboring slices, and similarly, iii) a dataset with five middle-most sagittal slices. In each dataset, individual slices were considered as independent samples, i.e., inputs for the classification method. The developed classification method yielded the best results when the input dataset comprised of three middle-most slices (Balanced Accuracy score (BACC) of 0.709 ± 0.011, Average Precision (AP) of 0.467 ± 0.025, and Area Under Receiver Operating Characteristic curve (ROC-AUC) of 0.740 ± 0.008). The baseline model trained using only the mid-sagittal slice for classification yielded the lowest classification scores (BACC of ...