A Two-stage Segmentation Neural Network for MICCAI FLARE22 Challenge

Jianping Zhang, Shuhui Jia

Keywords: Segmentation · Pseudo Supervision · Attention Block.

Abstract: Abdominal organ segmentation plays an important role in medical image processing. In this work, our goal is to segment thirteen organs of the abdomen in a semi-supervised way. We apply the attention block to the DMFNet, and propose a new Attention DMFNet for medical imaging, which can automatically learn and focus target structures of different shapes and sizes. The DMFNet is a highly efficient 3D CNN, which can realize real-time dense volume segmentation. It uses 3D multi fiber units composed of lightweight 3D convolution network to significantly reduce the computational cost. Models trained with attention blocks implicitly learn to suppress irrelevant regions in an input image while highlighting salient features useful for a specific task. Integrating attention blocks into the DMFNet can improve model sensitivity and prediction accuracy with minimal computational overhead. And we adopt a two-stage approach. In the first stage, the foreground containing all organs is segmented. In the second stage, thirteen organs are segmented on the basis of the first stage. We use labeled data to train the teacher model, use the teacher model to predict the unlabeled data, and take the segmentation result as pseudo labels for the following training. And then the data with true labels and the data with pseudo labels are used to train student models with the help of robust loss functions, namely beta cross-entropy, symmetric cross-entropy, and generalized cross-entropy. Finally, the trained student model is used to predict the data.