Active Contour with Adaptive Timestep for Brain Tumor Segmentation

Active contours with level set have been widely used across various image segmentation applications. Leakage of the active contour around weak edges is a major challenge, specially for medical image segmentation. One of the factors responsible for the leakage is a fixed time-step used in the discrete implementation of the active contour evolution. A smaller time-step results in a slow-steady evolution of the contour increasing the overall evolution time till its convergence. Whereas a larger time-step may result in skipping over weak edges resulting in erroneous segmentation results. Therefore an adaptive time-step scheme is proposed in this work. This scheme manages a larger time-step when contour is evolving in the homogeneous regions, while it updates to lower time-steps as the contour gets on reaching the boundary. It continuously adapts the time-step using a locally explored feature of the contour. This strategy is implemented on an edge-based active contour in this work. It is then evaluated on a MRI dataset having three kinds of tumors. The results show improvement in segmentation performance over the fixed time-step approach of the original edge-based active contour. Our results are also compared with a recent CNN-based segmentation on the same dataset. It is found that our method gives better tumor delineation results measured in terms of Dice Similarity Coefficient.