Developing an Automatic Cooperating Neural Networks and Image Standardization Approach for Segmentation of X-Ray Computed Tomography Images

Medical image segmentation is a common image analysis task that can impact diagnostic accuracy. Image quality varies depending on application, acquisition parameters and patient preparation. This requires developers’ data preparation efforts prior training. Training on an explicit data nature imposes limitations on generalization. This requires further data pre/post processing efforts for each dataset of different nature. Hence, it’s essential to embed a core component in the segmentation algorithm to automatically handle data processing tasks. In this work, we developed a fully-automated system that transforms computed tomography (CT) images into a standard format prior to segmentation learning and then inverse-transform segments to the original image format to preserve co-registration. Furthermore, we have developed a multi neural network segmentation module that incorporates three convolutional task-specific neural networks that utilize global (two networks) and local (one network) information, while making the segmentation based on agreement of contextual information using probability heatmaps. For baseline non-standard training and testing on a diverse liver dataset, there was no learning convergence and testing results performed relatively poorly. In contrast, after standardized training on the same dataset, all networks converged and yielded dice of +90%. Using the same standardized training, standard and non-standard testing were compared after applying single variable variation. Standardization consistently improved performance by 2.5%, 34.7% and 1.5% for field-of-view, slice-thickness and anatomy, respectively. We also found that global-based segmentation outperformed local-based, and network cooperation significantly outperformed individual networks with dice 95.20%. This offers a segmentation system with better generalization capabilities to serve variable clinical environments.