This paper presents a method to determine the stereoradiographic planes and anatomical vertebral landmarks giving the most reliable three-dimensional reconstructions of the thoracic and lumbar spine for clinical studies. The present investigation was limited to stereoradiographic setups with a normal vertical stereo base. Possible X-ray tube positions are thus corresponding to angles ranging from 0 (conventional posteroanterior radiograph) up to 30 degrees (dimension of the X-ray room). An X-ray phantom was used as a specimen from which three-dimensional reconstructions with the direct linear transformation (DLT) algorithm were obtained. Visibility of landmarks located on pedicles, end-plates, transverse and spinous processes was evaluated for the whole thoracic and lumbar spine (T1 to L5). Process landmarks were discarded because their poor visibility on radiographs produced inaccurate three-dimensional reconstructions. Considering the size, shape and orientation of vertebrae, an angle of 20 degrees between the posteroanterior horizontal position and the angled position of the X-ray tube gave optimal results. Landmarks located on pedicles and end-plates produced the most reliable three-dimensional reconstructions of the spine. Pedicles were found to be more reliable landmarks than end-plates. Validation of the technique with reconstructed steel beads reveals three-dimensional errors under 1.0 mm. Since vertebral landmarks were more difficult to identify on radiographs than steel beads, reconstruction results were compared with those obtained with a biplanar orthogonal setup. This shows that three-dimensional errors of 8.0 mm may be expected on actual reconstructions of the spine and errors as large as 15.0 mm may be present on poorly visible landmarks.