Abstract:
In modern medical diagnostics, computer vision and deep learning play an increasingly important role, especially in the analysis of complex 3D medical images. A significant obstacle to the implementation of modern deep learning algorithms in clinical practice are artifacts and inaccuracies of the primary classification by neural networks. In this paper, we systematized the main post-processing methods used in medical image segmentation tasks and reviewed related works on this topic. The aim of the study is to develop post-processing methods to eliminate segmentation errors associated with spatial incoherence and incorrect classification of 3D image voxels. In this paper, we propose a post-processing module for CT image segmentation results that effectively solves the problems of intersecting and nested pathologies. Three algorithms have been developed and implemented to eliminate fragments of false positive responses of the neural network. Experimental verification has shown that the proposed algorithms successfully provide unified coherent pathologies, which improves the quality of segmentation and simplifies subsequent analysis. The developed post-processing module can be integrated with the existing neural network framework for segmentation of medical images nnU-Net, which will contribute to improving the quality of diagnostics. The results of the study open up prospects for further development of post-processing methods in the field of medical imaging and can find wide application in systems for supporting medical decision-making.
Keywords:post-processing, computer vision, deep learning, image segmentation, medical diagnostics, probability maps.
Fulltext:
PDF file (538 kB)