Adaptive Refinement of Segmented Object Contour Based on the Brightness of Neighboring Pixels Using the Ensemble Method

Improving the accuracy of computer vision algorithms plays a significant role in the tasks of medical image segmentation. After all, determining the boundaries of objects is a difficult task when using medical images, and especially X-ray images. The use of X-ray images in segmentation tasks is a complex process, since these images themselves can have a sufficient amount of noise and artifacts. Classical segmentation methods face significant challenges when segmenting X-ray images where there are objects with fuzzy boundaries. To solve such tasks, it is suggested to use segmentation with the help of machine learning, and to increase the accuracy of determining the boundaries of objects, it is necessary to use adaptive approaches. This paper proposes a new method to improve the accuracy of X-ray image segmentation, which analyzes the neighboring pixels of each contour element and adaptively reshapes it if necessary, and then combines all predictions using an ensemble method, which improves the previous version of the contour. The method was able to demonstrate an improvement in the quality of image segmentation on three datasets with different complexity of structures. Improvements in object boundary accuracy were obtained for all three sets.