Automatic semantic segmentation of the lumbar spine: Clinical applicability in a multi-parametric and multi-center study on magnetic resonance images.

Significant difficulties in medical image segmentation include the high variability of images caused by their origin (multi-center), the acquisition protocols (multi-parametric), the variability of human anatomy, illness severity, the effect of age and gender, and notable other factors. This work addresses problems associated with the automatic semantic segmentation of lumbar spine magnetic resonance images using convolutional neural networks. We aimed to assign a class label to each pixel of an image, with classes defined by radiologists corresponding to structural elements such as vertebrae, intervertebral discs, nerves, blood vessels, and other tissues. The proposed network topologies represent variants of the U-Net architecture, and we used several complementary blocks to define the variants: three types of convolutional blocks, spatial attention models, deep supervision, and multilevel feature extractor. Here, we describe the topologies and analyze the results of the neural network designs that obtained the most accurate segmentation. Several proposed designs outperform the standard U-Net used as a baseline, primarily when used in ensembles, where the outputs of multiple neural networks are combined according to different strategies.

Appropriateness of lumbar spine magnetic resonance imaging in Spain.

OBJECTIVES: To determine the minimum percentage of lumbar spine magnetic resonance imaging (LSMRI) which are inappropriately prescribed in routine practice. METHODS: LSMRI performed prospectively on 602 patients in 12 Radiology Services across 6 regions in Spain, were classified as "appropriate", "uncertain" or "inappropriate" based on the indication criteria established by the National Institute for Clinical Excellence, the American College of Physicians and Radiology, and current evidence-based clinical guidelines. Studies on patients reporting at least one "red flag" were classified as "appropriate". A logistic regression model was developed to identify factors associated with a higher likelihood of inappropriate LSMRI, including gender, reporting of referred pain, health care setting (private/public), and specialty of prescribing physician. Before performing the LSMRI, the radiologists also assessed the appropriateness of the prescription. RESULTS: Eighty-eight percent of LSMRI were appropriate, 1.3% uncertain and 10.6% inappropriate. The agreement of radiologists' assessment with this classification was substantial (k=0.62). The odds that LSMRI prescriptions were inappropriate were higher for patients without referred pain [OR (CI 95%): 13.75 (6.72; 28.16)], seen in private practice [2.25 (1.20; 4.22)], by orthopedic surgeons, neurosurgeons or primary care physicians [2.50 (1.15; 5.56)]. CONCLUSION: Efficiency of LSMRI could be improved in routine practice, without worsening clinical outcomes.