This two-center retrospective study developed a two-stage deep learning (DL) model to automatically evaluate radiological features of facet joint osteoarthritis (FJOA) on lumbar CT scans. Researchers analyzed 13,223 facet joint CT images from 1,360 patients, dividing them into training, validation, and internal and external testing sets. The first stage, using nnU-Net, automatically detected and segmented facet joints with a Dice score of 0.81. The second stage used a ResNet-18 model to classify five FJOA features: joint space narrowing (JSN), osteophytes, hypertrophy, subchondral bone erosions, and subchondral cysts, with model attention visualized using Grad-CAM. On internal testing, the model reached accuracies of 89.8% for JSN, 79.6% for osteophytes, 65.5% for hypertrophy, 88.0% for subchondral bone erosions, and 82.8% for subchondral cysts. External testing showed similar performance, with accuracies between 56% and 89.8%. Overall, reliability was strong (Gwet κ up to 0.88). When two junior radiologists used the model, their diagnostic accuracy improved significantly (p < 0.05), and average reading time per image dropped from about 40 seconds to 15–20 seconds.