Sagittal views provide detailed and critical information about the anatomy and pathology in diagnosing and managing lumbar spine diseases. Radiologists comprehensively evaluate the spinal alignment, structural integrity, and health of bone and soft tissue elements in the sagittal view in diagnosing various spinal conditions and planning appropriate treatments. The first step in the automatic diagnosis is localizing the region of interest, typically the lumbar spine segments. To accomplish this, we present a modified U-Net (MU-Net) for the segmentation and localization of the lumbar spine from sagittal views of Magnetic Resonance Imaging (MRI) images and compare the performances with a series of state-of-the-art YOLOv implementations. We employ different techniques to augment data, address issues with limited training samples, and improve the generalization of deep models. We use the two YOLOv series for localization and MU-Net for segmentation. The MU-Net model achieves an accuracy of 98.97%, accompanied by a Mean Intersection Over Union (IoU) of 82.49%. For the localization, YOLOv8 yields a Precision, Recall, and Mean Average Precision (mAP) of 99.6%, 99.2%, and 99.4%, respectively.