Abstract
INTRODUCTION: Brain glioma segmentation is a critical task for medical diagnosis, monitoring, and treatment planning. DISCUSSION: Although deep learning-based fully convolutional neural networks have shown promising results in this field, their unstable segmentation quality remains a major concern. Moreover, they do not consider the unique genomic and basic data of brain glioma patients, which may lead to inaccurate diagnosis and treatment planning. METHODS: This study proposes a new model that overcomes this problem by improving the overall architecture and incorporating an innovative loss function. First, we employed DeepLabv3+ as the overall architecture of the model and RegNet as the image encoder. We designed an attribute encoder module to incorporate the patient's genomic and basic data and the image depth information into a 2D convolutional neural network, which was combined with the image encoder and atrous spatial pyramid pooling module to form the encoder module for addressing the multimodal fusion problem. In addition, the cross-entropy loss and Dice loss are implemented with linear weighting to solve the problem of sample imbalance. An innovative loss function is proposed to suppress specific size regions, thereby preventing the occurrence of segmentation errors of noise-like regions; hence, higher-stability segmentation results are obtained. Experiments were conducted on the Lower-Grade Glioma Segmentation Dataset, a widely used benchmark dataset for brain tumor segmentation. RESULTS: The proposed method achieved a Dice score of 94.36 and an intersection over union score of 91.83, thus outperforming other popular models.