Multi-Slice Dense-Sparse Learning for Efficient Liver and Tumor Segmentation

TL;DR

This paper introduces a novel multi-slice dense-sparse learning approach combined with a lightweight 2.5D nnU-Net architecture, significantly improving liver and tumor segmentation efficiency and accuracy in medical images.

Contribution

It proposes a new dense-sparse training flow and a lightweight 2.5D network architecture to enhance segmentation performance while reducing computational costs.

Findings

Outperforms existing methods on the LiTS dataset

Improves model efficiency with depthwise separable convolutions

Enhances segmentation accuracy through dense-sparse slice regularization

Abstract

Accurate automatic liver and tumor segmentation plays a vital role in treatment planning and disease monitoring. Recently, deep convolutional neural network (DCNNs) has obtained tremendous success in 2D and 3D medical image segmentation. However, 2D DCNNs cannot fully leverage the inter-slice information, while 3D DCNNs are computationally expensive and memory intensive. To address these issues, we first propose a novel dense-sparse training flow from a data perspective, in which, densely adjacent slices and sparsely adjacent slices are extracted as inputs for regularizing DCNNs, thereby improving the model performance. Moreover, we design a 2.5D light-weight nnU-Net from a network perspective, in which, depthwise separable convolutions are adopted to improve the efficiency. Extensive experiments on the LiTS dataset have demonstrated the superiority of the proposed method.