Improving the Random Walker algorithm for interactive 3D medical image segmentation using AI predictions

Abstract

The segmentation of anatomical structures in 3D medical images is crucial for various applications in the field of medical imaging. Fully automated methods often lack accuracy, while manual segmenting requires much time and effort from a user. Due to this, Active Learning approaches are being proposed to solve this by making the method interactive. This is done by iteratively segmenting the image based on user provided input, which can then be added to correct uncertain regions of intermediate segmentation results. We propose a method to improve a Random Walker (RW) algorithm that is used for interactive 3D medical image segmentation, by integrating an ensemble of predicted segmentations obtained by a previously trained Bayesian Deep Neural Network (BDNN). The predictions are used to improve the weight function used by the RW to indicate the similarity of neighbouring voxels. We evaluate our results by combining the weight function originally used by the RW, with two custom approaches, Mean Predictions and Unanimous Votes. Both are combined with the original RW's weights in the form of a weighted sum. Mean Predictions is based on only the mean of all predictions for each voxel, while Unanimous Votes only considers averages of either exactly 1 or 0. Lastly, we will propose a method to combine the original weight function with the mean predictions by summing together the image intensities with the mean predictions while keeping the variance normalized. All results are evaluated using both synthetic data and empirical data from a MICCAI dataset. Lastly, our method shows that the Adaptive Alpha Approach outperforms all other methods including the original RW in terms of average DICE Coeffients for the first two iterations.