A semantic segmentation model for lumbar MRI images using divergence loss

Automatic diagnosis of lumbar diseases is very important for improving diagnostic efficiency and optimizing the allocation of medical resources. Lumbar spinal stenosis (LSS) is a common disease of the lumbar spine that causes lower back pain and leg pain. Some geometric indicators of axial lumbar magnetic resonance imaging (MRI), such as the intervertebral disc area, the dural sac area, and the anteroposterior diameter, are important for diagnosis and treatment. An important prestep for the automatic measurement of these geometric indicators is semantic image segmentation. Traditional medical image semantic segmentation models generally use the pixel-wise loss function as the optimization objective, which produces illogical segmentation results during inferring and interferes with subsequent measurements of geometric indicators. We introduce Gaussian divergence loss (DVG-loss) and, combined with contour loss, propose a new loss function to optimize the segmentation model and achieve better results in the lumbar MRI image segmentation task. The improvement brought by the proposed loss function is mainly reflected in the geometrical appearance of segmentation results instead of the pixel-wise quantitative metrics. But we must make sure that the quantitative metrics won’t deteriorate. So first, we compare the proposed method with previous models quantitatively. And then, ablation studies are conducted on different loss functions and it is shown that our proposed loss function considerably reduces the irregular edges and isolated island regions caused by misclassification.