Median Filter Helps Lymph Node Segmentation in Deep Learning via PET/CT

Abstract

Pathological lymph node segmentation plays a vital role in clinical practice. Yet it is still a challenging problem owing to low contrast to surrounding structures on images. In this paper, we investigate the problem whether the classical median filter helps lymph node segmentation in deep learning via PET/CT. Specifically, we design a median filter layer and integrate it into two types of deep convolution neural networks: SegNet and DeepLabv3+, which takes the encoder and decoder structure that has the advantage to segment objects in a multi-scale way. Meanwhile, we adopt three various objective functions, which are cross entropy loss, generalized Dice loss and focal loss, to study which is the best choice for pathological lymph node segmentation with median filter. Four-fold cross validation has been done on 63 volumes containing 214 malignant lymph nodes, and the experiments demonstrate that median filter could help improve the lymph segmentation performance with cross entropy as loss function, which has 3% and 2% improvements with SegNet and 4% and 3% improvements with DeepLabv3+ in terms of Sensitivity and Dice Similarity Coefficient (DSC).