P28. Deep learning-based detection of lumbar spinal canal stenosis using convolutional neural networks

Lumbar spinal canal stenosis (LSCS) is the most common spinal degenerative disease in elderly people and usually first seen by primary care physicians or orthopedic surgeons who are not spine surgery specialists. Magnetic resonance imaging (MRI) is useful in the diagnosis of LSCS, but the equipment is often not available or difficult to read. LCSC patients with progressive neurologic deficits have difficulty with recovery if surgical treatment is delayed. So, early diagnosis and determination of appropriate surgical indications are crucial in the treatment of LCSC. Convolutional neural networks (CNNs), a type of deep learning, offers significant advantages for image recognition and classification, and work well with radiographs, which can be easily taken at any facility. Our purpose was to develop an algorithm to diagnose the presence or absence of LSCS requiring surgery from plain radiographs using CNNs. This study is a cross-sectional study. One hundred patients who underwent the surgery for LSCS including degenerative spondylolisthesis from January 2022 to May 2022 at a single institution were enrolled. In annotation 1, the area under the curve (AUC) computed from the receiver operating characteristic (ROC) curve, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, positive likelihood ratio (PLR), and negative likelihood ratio (NLR) were calculated. In annotation 2, correlation coefficients were used. Four intervertebral levels from L1/2 to L4/5 were extracted as region of interest from lateral plain lumbar spine radiographs and totally 400 images were obtained. Based on the date of surgery, the 300 images derived from the first 75 cases were used for internal validation and 100 images from the second 25 cases for external validation. In annotation 1, binary classification of operative and nonoperative levels was used, and in annotation 2, the spinal canal area rate was calculated by dividing each disc level area measured on the MRI axial image by L1/2 level area. For internal validation, 300 images were divided into each 5 datasets on per-patient basis and 5-fold cross-validation was performed. Five trained models were registered in the external validation prediction performance. Grad-CAM was used to visualize area with the high features extracted by CNNs. In internal validation, the range of AUC and accuracy were 0.80 to 0.96 and 75% to 93% for the annotation 1 and correlation coefficients of 0.60 to 0