Low back pain is a widespread condition and a major contributor to disability worldwide, with spondylolisthesis being one of its associated pathologies. This study evaluates the performance of four convolutional neural network (CNN) architectures—DenseNet201, ResNet152, VGG16, and Xception—in the automatic detection of spondylolisthesis from lumbar X-rays. 10,616 radiographic images from the BUU-LSPINE dataset were used, including anteroposterior (AP) and lateral (LA) projections from 5,308 patients. Three training scenarios were considered: full dataset, LA-only, and AP-only. ResNet152 consistently achieved the highest overall accuracy (88.34%) and AUC (93.64%), while DenseNet201 obtained the highest precision (97.01%) and specificity (98.12%), indicating strong performance in minimizing false positives. VGG16 showed balanced results across metrics, with the highest recall (79.44%) on AP images, whereas Xception underperformed across all evaluation criteria, indicating limited generalization capability. Although the findings are promising, limitations such as a lack of external validation, the absence of advanced preprocessing, and potential demographic bias must be addressed. Future studies should focus on improving generalizability, incorporating explainability tools, and validating models in diverse clinical environments to support reliable deployment in diagnostic workflows.