Using a convolutional neural network model to derive imaging landmarks for lumbar spine numbering on axial magnetic resonance images

Understanding the axial lumbar spine anatomy, including knowledge of the relationship between the lumbar spine level and other paraspinal structures, is important for diagnosing and treating diseases. The purpose of this study was to validate the accuracy of a convolutional neural network (CNN) model in lumbar spine level numbering on axial magnetic resonance (MR) images and to find the appropriate anatomic landmarks for numbering using a class activation map (CAM). A total of 6055 axial MR images of the lumbar spine from the L1‐2 to L5‐S1 disc levels were obtained to train and validate the CNN model. MR images were acquired using three 3‐Tesla machines. The algorithm was developed with three models, and the best‐performing model was selected. The external validation set (n = 493) was obtained from other institutions using various machines. The accuracy of the numbering was analyzed using a confusion matrix and receiver operating characteristic curves. The CAMs were reviewed, and the identified anatomic structures were investigated. A reader study was performed by three radiologists, and their accuracy was compared with that of the model. The overall accuracy of the best‐performing model for lumbar spine numbering was 0.98 on internal validation and 0.95 on external validation. For the CAM review, mappings concentrated on both paraspinal areas, including the kidney, back muscles, and ilium according to the level. Top‐1 and top‐2 accuracies of the reviewers ranged between 0.56–0.75, and 0.84–0.93, respectively. After reviewing the CAMs, the accuracy increased to 0.75–0.78 and 0.93–0.98, respectively. A CNN model can accurately determine the level of the lumbar spine on axial MR images, and the configuration of muscles can be used to determine the lumbar level.