Lighter U-net for segmenting white matter hyperintensities in MR images

Jun Zhuang, Mingchen Gao, Mohammad Al Hasan

TL;DR

This paper introduces "Lighter U-Net," a novel deep learning architecture designed for efficient and accurate segmentation of white matter hyperintensities (WMH) in MRI scans. By integrating dense connections and dilated convolutions into the U-Net framework, the model significantly reduces parameters and computational cost while maintaining or improving segmentation performance compared to standard U-Net.

What Problem Does This Paper Address?

White matter hyperintensities (WMH) are markers of small vessel diseases and are crucial indicators in clinical research and practice. Manually segmenting these lesions is time-consuming and prone to inter-observer variability. While automated methods, particularly U-Net, have shown promise, they often suffer from a loss of detailed information as the network deepens. Furthermore, existing U-Net architectures can be computationally expensive, hindering real-time applications, especially on cloud platforms. This paper aims to address these limitations by proposing a more efficient and effective U-Net variant.

Methods

The proposed "Lighter U-Net" builds upon the standard U-Net architecture by incorporating two key innovations:

  • Dense Connections: Inspired by DenseNet, each layer in the Lighter U-Net receives feature maps from all preceding layers as input. This promotes feature reuse, helps alleviate the vanishing-gradient problem, and allows for more efficient parameter utilization.
  • Dilated Convolutions: To counteract the resolution loss that occurs in deeper layers of U-Net, dilated convolutions are employed. These convolutions expand the receptive field without increasing the number of parameters or downsampling the feature maps, thus preserving finer details crucial for accurate segmentation.

The Lighter U-Net architecture integrates these concepts, particularly in the decoder path, to enhance feature representation and reduce computational overhead. The authors experimented with different configurations, ultimately settling on a 10-layer network with specific dense connection parameters and a dilated convolution rate of 2.

Datasets

The study utilized a dataset acquired from three different hospitals in the Netherlands and Singapore. Each patient's data included T1-weighted images, FLAIR images, and manual WMH labels. The dataset comprised 60 patients in total, with images originating from three different scanners (Philips Achieva, Siemens TrioTim, and GE Signa HDxt). For training and evaluation, a five-fold cross-validation strategy was employed, with 80% of the data used for training, 20% for validation, and the remaining for testing within each fold. The paper notes that no data augmentation techniques were used during preprocessing.

Key Insights and Ideas

  • Efficiency Through Feature Reuse: Dense connections are a powerful mechanism for improving model efficiency by allowing layers to leverage features from earlier stages, reducing the need for redundant computations and parameters.
  • Preserving Detail with Dilated Convolutions: Dilated convolutions offer a clever way to increase the receptive field of a network without sacrificing spatial resolution, which is critical for segmenting intricate structures like WMH.
  • Trade-off Between Performance and Size: The research highlights that while deeper networks or more filters can sometimes improve accuracy, they come at a significant computational cost. A lighter architecture can achieve comparable or even superior performance in terms of efficiency.
  • Preprocessing Matters: The study found that padding was a more effective preprocessing strategy than cropping for achieving higher segmentation accuracy, underscoring the importance of data preparation.
  • Batch Normalization Placement: The position of batch normalization layers can influence performance, with placement before convolutional layers yielding better results in this architecture.

Key Results

The Lighter U-Net demonstrated significant improvements over the standard U-Net:

  • Segmentation Accuracy: The proposed model achieved a testing Dice similarity coefficient (DSC) of 79.28%, outperforming the standard U-Net's 71.03%.
  • Parameter Reduction: The Lighter U-Net uses only 17% of the parameters of a standard U-Net, making it significantly lighter.
  • Speed Improvement: The model achieved a 26% reduction in running time compared to the standard U-Net.
  • Weight Reduction: The model's weight size was reduced by 83%.
  • Ablation Study: The paper showed that dense connections were crucial for achieving good performance with fewer filters, and dilated convolutions further improved accuracy and reduced model weight.

Limitations and Future Work

The authors acknowledge two main limitations:

  • Memory Consumption: Dense concatenation, while beneficial for performance, can lead to increased memory usage.
  • Further Compression: The current Lighter U-Net, while lighter than standard U-Net, can still be further compressed to achieve even faster real-time segmentation.

Future work will focus on addressing these limitations by exploring more aggressive model compression techniques to enable real-time processing on resource-constrained devices.

Summary updated

Abstract

White matter hyperintensities (WMH) is one of main consequences of small vessel diseases. Automated WMH segmentation techniques play an important role in clinical research and practice. U-Net has been demonstrated to yield the best precise segmentation results so far. However, sometimes it losses more detailed information as network goes deeper. In addition, it usually depends on data augmentation or a large number of filters. Large filters increase the complexity of model, which may be an obstacle for real-time segmentation on cloud computing. To solve these two issues, a new architecture, Lighter U-Net is proposed to reinforce feature use, to reduce the number of parameters as well as to retain sufficient receptive fields without losing resolution. The extensive experiments suggest that the proposed network achieves comparable performance as the state-of-the-art methods by only using 17% parameters of standard U-Net.
BibTeX
@inproceedings{zhuang2019lighter,
  author = {Jun Zhuang and Mingchen Gao and Mohammad Al Hasan},
  title = {{Lighter U-net for segmenting white matter hyperintensities in MR images}},
  booktitle = {MobiQuitous},
  year = {2019},
  doi = {10.1145/3360774.3368203},
}
Figure