The prevalence of vertebral fractures has increased, especially those caused by osteoporotic compression, becoming a sensitive public health issue. The diagnosis of such injuries can be performed in different ways, many of which have limitations regarding cost, availability, and time to analyze the results. Among the possible evaluation options is the use of conventional radiographs (X-rays), a mechanism that has been refined lately. This raises the possibility of applying machine learning techniques as a promising alternative to support and accelerate clinical diagnosis. To address this problem, this study proposes the use of deep learning-based computer vision for vertebral segmentation in radiographic images, considering it an important step for some lesion classification mechanisms. In this case, lateral radiographic images of the lumbar spine were used in conjunction with the respective annotations provided by the BUULSPINE dataset. From this, a segmentation scheme was structured based on the U-Net model. Due to the lack of precise annotations, the use of two distinct sets of masks for training was proposed: one based on quadrilaterals generated from the annotated vertices and another using masks generated semi-automatically through the Segment Anything Model. Next, the performance of the proposed approaches was evaluated using metrics commonly adopted in image segmentation activities: Dice-Sørensen and Jaccard coefficients. The results obtained reinforce the viability of the U-Net model in the application of the activity in question and indicate that the two strategies adopted for mask creation result in similar performances, with average test performances around 0.974 (Dice) and 0.949 (Jaccard).