Deep Active Contour-Based Capsule Network for Medical Image Segmentation

Abstract

Listen

Translate article

Medical Image Segmentation (MIS) has important ramifications for the whole of medical diagnostics. Despite the unparalleled success that various Deep Learning (DL) techniques had in image segmentation, there are setbacks with them. Generally, these setbacks arise from two of the most profound indicators of performance in DL, namely network architecture and loss function, both of which are being dealt with in the proposed method to improve the dice score on the task of MIS. To deal with the architecture part of our work, we have used Capsules that have proved to preserving more image details and indicating the spatial relationships found locally and globally, facilitating the detection of global features from the local features extracted, and hence we propose a modified capsules network for the proposed segmentation task. The loss function used in the proposed work draws inspiration from the active contours model that tends to incorporate external forces, regional information and other functions of choices in the working of segmentation extraction and treats this task as a continuous curve evolution and energy minimization problem. In our experimentation, we have used a brain tumor segmentation dataset for the performance comparison using Dice-score and mean-IoU as performance metrics and the results obtained are very optimistic.