This study evaluated whether lightweight deep learning (DL) models can classify lumbar foraminal stenosis (LFS) severity from sagittal CT scans, providing an alternative when MRI is unavailable. In a retrospective, single-center study at Wenzhou Medical University (China), researchers analyzed 868 sagittal CT images from 177 patients collected between 2016–2025, all of whom also underwent MRI within 7 days. LFS grades were assigned using the Lee system by two surgeons with adjudication by a third. Two DL models were developed (EfficientNet-B0 and MobileNetV3-Large-100) and were trained and tested against a senior (18 years experience) and junior (5 years experience) spine surgeon using Faster R-CNN/ResNet-50 region of interest (ROI) detection. On the independent test set (102 images, 21 patients), accuracies were 82.35% (DL1) and 80.39% (DL2), nearly identical to the senior surgeon (83.33%) and much higher than the junior (62.75%). Agreement with the senior surgeon was strong (k = 0.815, DL1; k = 0.799, DL2). For binary detection of nerve compression, accuracies were 87–89% (models) vs. 91% (senior). Limitations include single-center sampling, modest dataset size, and reliance on subjective grading. No external validation or subgroup analysis was performed.