Clinical Learning Order-Guided Deep Neural Network for Brain Tumor Segmentation

Abstract

In this paper, we present an innovative multi-label local-to-global learning segmentation (MLLGL-Seg) neural network model for brain tumor segmentation. Our framework is grounded on a clinical learning order from the whole tumor (WT) to the tumor core (TC) and then to the enhanced tumor (ET). Thus, we first propose a multi-label segmentation network to embed the clinical learning sequence of WT-TC-ET during training. We then introduce a local-to-global learning algorithm and integrate it into multi-label to construct the MLLGL-Seg model. In addition, considering the hierarchical structure of the output tumor region, we further perform hierarchical consistency transformation on the network output to ensure that it complies with hierarchical constraints. The novelty of this paper lies in the proposal of the MLLGL-Seg by introducing curriculum learning based on category space into segmentation, in the construction of a learning sequence based on boundary difficulty and category similarity based on clinical experience, and in the device of learning sequences based on anti-similarity and label noise. Experimental results show that the proposed method achieves one of the most competitive results in brain tumor segmentation accuracy on three publicly available datasets.