DMC-Net: Lightweight Dynamic Multi-Scale and Multi-Resolution Convolution Network for Pancreas Segmentation in CT Images

TL;DR

The proposed DMSC and DMRC modules are designed as lightweight drop-in replacements for single convolutions, and they can be easily integrated into general CNN architectures for end-to-end training.

Abstract

Convolutional neural networks (CNNs) have shown great effectiveness in medical image segmentation. However, they may be limited in modeling large inter-subject variations in organ shapes and sizes and exploiting global long-range contextual information. This is because CNNs typically employ convolutions with fixed-sized local receptive fields and lack the mechanisms to utilize global information. To address these limitations, we developed Dynamic Multi-Resolution Convolution (DMRC) and Dynamic Multi-Scale Convolution (DMSC) modules. Both modules enhance the representation capabilities of single convolutions to capture varying scaled features and global contextual information. This is achieved in the DMRC module by employing a convolutional filter on images with different resolutions and subsequently utilizing dynamic mechanisms to model global inter-dependencies between features. In contrast, the DMSC module extracts features at different scales by employing convolutions with different kernel sizes and utilizing dynamic mechanisms to extract global contextual information. The utilization of convolutions with different kernel sizes in the DMSC module may increase computational complexity. To lessen this burden, we propose to use a lightweight design for convolution layers with a large kernel size. Thus, DMSC and DMRC modules are designed as lightweight drop-in replacements for single convolutions, and they can be easily integrated into general CNN architectures for end-to-end training. The segmentation network was proposed by incorporating our DMSC and DMRC modules into a standard U-Net architecture, termed Dynamic Multi-scale and Multi-resolution Convolution network (DMC-Net). The results demonstrate that our proposed DMSC and DMRC can enhance the representation capabilities of single convolutions and improve segmentation accuracy.

Figures (4)

• Figure 1: The architecture of 2D DMC-Net. It is designed as a U-shaped network, consisting of a 5-stage encoder, a 5-stage decoder, and a bottleneck. In each stage, the basic block consists of a 2D DMSC module and a 2D DMRC module. A $1\times1$ convolutional layer is utilized to project input features to $32$ channels in the encoder, and another $1\times1$ convolutional layer is utilized to make pixel-wise predictions in the decoder.
• Figure 2: The architecture of 3D DMC-Net. It is designed as a U-shaped network, consisting of a 4-stage encoder, a 4-stage decoder, and a bottleneck. In each stage, the basic block consists of a 3D DMSC module and a 3D DMRC module. A $1\times1\times1$ convolutional layer is utilized to project input features to $32$ channels in the encoder, and another $1\times1\times1$ convolutional layer is utilized to make voxel-wise predictions in the decoder.
• Figure 3: Visualizations of the pancreas segmentation results on the NIH TCIA-Pancreas dataset. The pancreas is marked in green. 2D DMC-Net shows better segmentation quality than 2D U-Net, and 3D DMC-Net shows better segmentation quality than 3D U-Net.
• Figure 4: Visualizations of the pancreas and pancreatic mass segmentation results on the MSD-Pancreas dataset. The pancreas is marked in green, and the tumor is marked in red. 2D DMC-Net shows better segmentation quality than 2D U-Net, and 3D DMC-Net shows better segmentation quality than 3D U-Net.