Focal Loss Function based DeepLabv3+ for Pathological Lymph Node Segmentation on PET/CT

Abstract

Pathological lymph node segmentation plays an important role in clinical practice. Yet it is still a challenging problem owing to low contrast to surrounding structures. In this paper, we take a deep learning based approach for pathological lymph node segmentation task. Semantic segmentation architecture, DeepLabv3+, which has the advantage to segment objects in a multi-scale way, is adopted in this paper. Meanwhile, the focal loss function, which originally applied in object detection task to deal with the imbalance class number, is integrated into DeepLabv3+ architecture for the imbalance of voxel class between pathological lymph nodes and background. Compared to the cross entropy loss function and dice function, the focal loss function can improve the segmentation performance in terms of sensitivity and dice in the DeepLabv3+ segmentation architecture. Four-fold cross validation has been done on 63 volumes containing 214 malignant lymph nodes and the mean sensitivity of 87% and average Dice score of 75% are obtained for pathological lymph node segmentation.