PaperlayerView exact revision on arXiv

Source-linked AI summary

DCSAU-Net: A Deeper and More Compact Split-Attention U-Net for Medical Image Segmentation

Qing Xu, Zhicheng Ma, Na HE, Wenting Duan

arXiv:2202.00972v2eess.IVcs.AIcs.CVcs.LG

TL;DR

Existing medical image segmentation networks may inadequately capture features from different depths and face increasingly challenging images. DCSAU-Net combines primary feature conservation with compact split-attention, and it achieves higher mIoU and F1 scores than other SOTA methods across four datasets.

  • Problem

    U-Net’s equal downsampling design and simply stacked convolutions may not extract sufficient features from different depths in increasingly complex medical images.

  • Method

    DCSAU-Net combines a primary feature conservation strategy for low-level information with a compact split-attention block for multi-scale channel representation.

  • Results

    DCSAU-Net achieves higher mIoU and F1 scores than other SOTA segmentation methods across four medical image segmentation datasets.

  • Takeaways & Limitations

    The architecture demonstrates strong segmentation performance on challenging medical images while using a compact design suitable for machines with limited memory.

Abstract

from arXiv · show

Deep learning architecture with convolutional neural network (CNN) achieves outstanding success in the field of computer vision. Where U-Net, an encoder-decoder architecture structured by CNN, makes a great breakthrough in biomedical image segmentation and has been applied in a wide range of practical scenarios. However, the equal design of every downsampling layer in the encoder part and simply stacked convolutions do not allow U-Net to extract sufficient information of features from different depths. The increasing complexity of medical images brings new challenges to the existing methods. In this paper, we propose a deeper and more compact split-attention u-shape network (DCSAU-Net), which efficiently utilises low-level and high-level semantic information based on two novel frameworks: primary feature conservation and compact split-attention block. We evaluate the proposed model on CVC-ClinicDB, 2018 Data Science Bowl, ISIC-2018 and SegPC-2021 datasets. As a result, DCSAU-Net displays better performance than other state-of-the-art (SOTA) methods in terms of the mean Intersection over Union (mIoU) and F1-socre. More significantly, the proposed model demonstrates excellent segmentation performance on challenging images. The code for our work and more technical details can be found at https://github.com/xq141839/DCSAU-Net.

1. Introduction

Medical image segmentation supports diagnosis but remains difficult on challenging images, while existing encoder-decoder models can lose high-resolution information. DCSAU-Net addresses these limitations with PFC and CSA modules for richer feature extraction.

  • Medical image analysis supports diagnosis and treatment, but manual analysis can be time-consuming, subjective, and error-prone.
  • Challenging images with low foreground-background contrast and complex tissue backgrounds require robust, generic segmentation models.
  • CNN encoder-decoder models use skip connections to combine low- and high-level semantics, but improved architectures may overlook high-resolution information at the encoder head.
  • DCSAU-Net combines primary feature conservation with a compact split-attention block to improve feature extraction across depths.
  • Across four medical image segmentation datasets, DCSAU-Net outperforms other SOTA methods on mIoU and F1 score.

2. Related Work

Prior work builds on U-Net’s encoder-decoder and skip-connection design, while attention and depthwise separable convolutions improve feature selection and efficiency. These approaches motivate DCSAU-Net’s compact, multi-scale design.

  • U-Net combines encoder features with decoder upsampling to recover resolution and achieve precise localization in medical images.
  • ResUNet++ extends u-shaped segmentation with residual connections, squeeze-and-excitation, ASPP, and attention mechanisms.
  • Channel and spatial attention mechanisms select informative feature channels or spatial locations for medical image segmentation.
  • Depthwise separable convolution reduces parameters and computation, then uses pointwise convolution to combine channel information into new feature maps.
  • Figure 1 compares PFC with U-Net, Stem block, and ResUNet++ designs for extracting low-level semantic information.

3. Method

DCSAU-Net preserves primary low-level features with PFC and extracts multi-scale representations with the CSA block. Residual connections and channel attention combine information across receptive fields.

  • Primary Feature Conservation: PFC uses residual depthwise separable convolutions and large kernels to preserve low-level features while limiting parameter and computation growth.
  • Primary Feature Conservation: Stacked convolutional layers can increase parameters and computation rapidly, while fixed receptive fields are inadequate for lesions of different sizes.
  • Compact Split-Attention block: CSA maintains two feature groups and assigns different convolutional depths to obtain feature maps with multiple receptive-field scales.
  • Compact Split-Attention block: CSA fuses one group with the other through a 3×3 convolution, combining semantic information to extract global and local features.
  • Compact Split-Attention block: Global average pooling produces channel-wise statistics, and channel-wise soft attention weights the split feature representations.
  • Compact Split-Attention block: The CSA block uses a residual skip connection, applying an additional transformation when input and output feature-map shapes differ.

4. Experiments and Results

The experiments evaluate DCSAU-Net across four publicly available medical image segmentation datasets covering polyp, nuclei, lesion-boundary, and multi-class plasma-cell segmentation. These datasets are linked to clinical diagnosis.

  • DCSAU-Net is evaluated on four publicly available medical image segmentation datasets.
  • The datasets cover polyp segmentation, nuclei segmentation, lesion-boundary segmentation, and multi-class segmentation.
  • CVC-ClinicDB supports polyp segmentation and served as the training database for the MICCAI 2015 Automatic Polyp Detection Challenge.
  • The datasets are related to clinical diagnosis, making their segmentation results potentially significant for patients.

4.2. Evaluation Metrics

The experiments use mIoU, Accuracy, Recall, Precision, and F1-score to evaluate medical image segmentation models, with mIoU highlighted as a common competition metric.

  • The evaluation calculates mIoU, Accuracy, Recall, Precision, and F1-score for model comparison.

4.3. Data Augmentation

Data augmentation addresses overfitting caused by limited medical-image training samples by increasing sample diversity and improving generalisation.

  • Limited and costly medical-image samples make segmentation models prone to overfitting.
  • Training randomly applies horizontal flips, rotations, and cutout with probability 0.25.

4.4. Implementation Details

All models are trained under a shared PyTorch setup, using common optimization settings and consistent dataset splits across the four experiments.

  • Experiments use PyTorch 1.10.0 on an NVIDIA V100 GPU with Dice loss and Adam at learning rate 1e-4.Batch size is 16 and training lasts 200 epochs.
  • All models use the same train, validation, and test datasets across the four experiments.Other SOTA models use default parameters; TransUNet and LeViT-UNet use pretrained ViT models, while the remaining models train from scratch.

4.5. Results

DCSAU-Net is evaluated against SOTA methods across four biomedical segmentation datasets and reports strong results on mIoU, F1-score, and other metrics.

  • CVC-ClinicDB: DCSAU-Net achieves F1-score 0.916 and mIoU 0.861 on CVC-ClinicDB.These values outperform DoubleU-Net by 2.0% in F1-score and 2.5% in mIoU.
  • SegPC-2021: DCSAU-Net achieves mIoU 0.8048 on SegPC-2021, rising 3.6% over Unet++.Its F1-score is 2.8% higher than DoubleUNet.
  • 2018 Data Science Bowl: DCSAU-Net achieves F1-score 0.914 and mIoU 0.850 on the 2018 Data Science Bowl dataset.The F1-score is 1.9% higher than TransUNet and the mIoU is 2.5% higher than UNet3+.
  • ISIC-2018: On ISIC-2018, DCSAU-Net improves mIoU by 2.4% over LeViT-UNet and F1-score by 1.8% over UNet3+.It also reports recall 0.922 and accuracy 0.960, better than other baseline methods.

4.6. Ablation Study

The ablation study evaluates PFC and CSA separately and together across four datasets, showing that both modules improve segmentation over U-Net configurations.

  • Ablation design: Table 6 evaluates four DCSAU-Net configurations using parameter count, FLOPs, FPS, mIoU, and F1-score across all four datasets.The study isolates module effectiveness and examines computational characteristics.
  • PFC strategy: 1.9% to 2.2% mIoU improvements result from adding PFC to U-Net across the four datasets.The gains are 1.9% on CVC-ClinicDB, 1.4% on SegPC-2021, 2.2% on the 2018 Data Science Bowl, and 1.9% on ISIC 2018.
  • CSA block: 1.5% to 3.1% mIoU improvements result from adding CSA to U-Net across the four datasets.The gains are 3.1% on CVC-ClinicDB, 1.5% on SegPC-2021, 3.1% on the 2018 Data Science Bowl, and 2.8% on ISIC 2018.
  • Module comparison: The CSA block has a more significant impact than PFC on mIoU in the reported U-Net comparisons.The comparison is based on the module-specific mIoU improvements in Table 6.
  • Combined configuration: 0.6% to 3.5% F1-score and 1.1% to 3.3% mIoU gains arise when both modules are combined versus U-Net with only PFC or CSA.The combined configuration is U-Net + PFC + CSA.

5. Discussion

The discussion frames DCSAU-Net as a response to lost primary features and explains how PFC and CSA support efficient, multi-scale feature representation. Qualitative and efficiency results indicate suitability for challenging images and memory-limited deployment.

  • Training results: The first 20 test epochs are presented for four medical image segmentation tasks.The figure provides an early-epoch view of test results.
  • PFC strategy: DCSAU-Net retains primary image features with depthwise separable convolutions and a larger kernel in its PFC strategy.The discussion notes that a 7x7 kernel was used for the best reported performance, despite smaller kernels reducing parameters and computation.
  • CSA block: CSA enhances cross-channel connectivity and multi-scale feature representation through attention and feature-map combination.Its branches use different convolution counts and receptive field sizes.
  • Deployment considerations: U-Net has shorter inference time, but DCSAU-Net uses a tiny number of parameters and acceptable inference time for limited-memory machines.The comparison concerns equal output feature channels.
  • Qualitative results: Qualitative results show DCSAU-Net captures more proper foreground information than other SOTA methods on challenging low-quality images.The cited examples include incomplete staining and obscurity.

6. Conclusion

The conclusion presents DCSAU-Net as an encoder-decoder architecture combining PFC and CSA for medical image segmentation. Across four datasets, it reports higher F1-score and mIoU than other SOTA models, especially for multi-class and complex images.

  • Architecture: DCSAU-Net combines PFC for primary-feature preservation with CSA for attention-based multi-receptive-field feature combination.CSA splits input feature maps into two groups with different convolution counts.
  • Results: Higher F1-score and mIoU scores are reported than other SOTA models across four medical image segmentation datasets.The conclusion states that performance is especially better on multi-class segmentation and complex images.
  • Future work: Future work will optimize DCSAU-Net for improved performance and applicability to more medical image segmentation tasks.This defines the stated scope of the authors’ future direction.
Loading 2202.00972v2…