Improving the segmentation of the vertebrae using a multi-stage machine learning algorithm

The health of the spine is an integral part of human health because the spine itself plays one of the key roles in human health, and diseases such as osteoporosis, vertebral injuries, herniated intervertebral discs, and other diseases can not only complicate a person's life but also have serious consequences. The use of X-ray images to diagnose spinal diseases plays a key role in diagnosis. Diagnosis of diseases with the help of X-rays is the most popular and cheapest option for patients to detect pathologies and diseases. The subjects of this article are algorithms for the segmentation of X-ray images of various qualities. The aim is to research the possibility of improving segmentation of vertebrae: Th8, Th9, Th10, Th11 using a multi-stage method of segmentation of the spine using machine learning to improve the accuracy of automation of vertebrae segmentation. Task: train a neural network that will segment the incoming X-ray image and produce a mask of the area of four vertebrae at the output; train a neural network that will segment each vertebra in the area found at the previous stage; cut out a section with one vertebra and train a neural network that will segment it; create an algorithm that, based on three previously trained neural networks, will segment vertebrae on an X-ray image. The following methods were used: a multi-stage approach using machine learning. The following results were obtained: thanks to segmentation in several stages, it was possible to reduce the region of interest, thereby removing unnecessary background when using segmentation. Using this algorithm for 48 vertebrae, an average improvement in segmentation accuracy of 4.83% was obtained. Conclusions. In this research, a multi-stage algorithm was proposed, and an improvement in the accuracy of segmentation of X-ray images in the lateral projection, namely the accuracy of all four vertebrae: Th8, Th9, Th10, Th11 - was obtained. The results demonstrate that the use of this method gives a better result than the usual segmentation of the input image.