Source-linked AI summary
CMU-Net: A Strong ConvMixer-based Medical Ultrasound Image Segmentation Network
Fenghe Tang, Lingtao Wang, Chunping Ning, Min Xian, Jianrui Ding
TL;DR
Medical ultrasound segmentation is difficult because lesions vary greatly and images often have low contrast, speckle noise, and artifacts. CMU-Net combines ConvMixer-based global context extraction with multi-scale attention gates, outperforming six comparison methods on breast and thyroid ultrasound datasets.
Problem
Medical ultrasound segmentation is challenging because lesion sizes, shapes, and textures vary, while images often have low contrast, speckle noise, and shadow artifacts.
Method
CMU-Net combines a ConvMixer module for global context information with multi-scale attention gates that emphasize valuable features during skip-connection transfer.
Results
CMU-Net achieves better segmentation performance than six comparison approaches, including a 2.89% IoU and 2.00% F1 improvement on the BUSI dataset.
Takeaways & Limitations
CMU-Net achieves state-of-the-art performance on two ultrasound datasets by combining global context modeling with efficient feature transfer.
Abstract
from arXiv · showhide
U-Net and its extensions have achieved great success in medical image segmentation. However, due to the inherent local characteristics of ordinary convolution operations, U-Net encoder cannot effectively extract global context information. In addition, simple skip connections cannot capture salient features. In this work, we propose a fully convolutional segmentation network (CMU-Net) which incorporates hybrid convolutions and multi-scale attention gate. The ConvMixer module extracts global context information by mixing features at distant spatial locations. Moreover, the multi-scale attention gate emphasizes valuable features and achieves efficient skip connections. We evaluate the proposed method using both breast ultrasound datasets and a thyroid ultrasound image dataset; and CMU-Net achieves average Intersection over Union (IoU) values of 73.27% and 84.75%, and F1 scores of 84.81% and 91.71%. The code is available at https://github.com/FengheTan9/CMU-Net.
1. INTRODUCTION
Medical ultrasound segmentation is difficult because lesions vary greatly and images often have low contrast, speckle noise, and shadow artifacts. CMU-Net addresses U-Net’s limited global context and basic skip connections with ConvMixer and multi-scale attention gates.
- Motivation: Ultrasound is non-invasive, non-radiative, cost effective, and real-time, but manual labeling is laborious and sensitive to radiologists’ experience and mental state.
- Motivation: Ultrasound lesions vary greatly in size, shape, and texture, while low contrast, speckle noise, and shadow artifacts make segmentation challenging.
- Motivation: Ordinary convolution locality limits U-Net’s ability to extract global information, motivating Transformer- and ConvMixer-based alternatives.ConvMixer mixes remote spatial locations with large convolutional kernels and has lower computational overhead than self-attention according to the cited discussion.
- Contribution: CMU-Net combines a ConvMixer module for global context with multi-scale attention gates for suppressing irrelevant and enhancing valuable features.The design is a fully convolutional segmentation network built on a U-shaped architecture.
- Contribution: CMU-Net’s contributions include fully convolutional design, attention-gated skip connections, and improved segmentation performance on breast and thyroid ultrasound images.
2. METHOD
CMU-Net uses a five-level encoder-decoder that combines ordinary convolutions, a ConvMixer block, bilinear upsampling, and multi-scale attention-gated skip connections. The design preserves spatial resolution in ConvMixer features while adaptively selecting encoder features across receptive fields.
- Encoder stage: The encoder extracts high-level semantics with ordinary convolutions, then feeds features into ConvMixer to obtain mixed spatial and location information.The encoder has five levels, each using two convolution blocks and max pooling, with ConvMixer at the final level.
- Encoder stage: ConvMixer layers combine depthwise k×k and pointwise 1×1 convolutions with GELU activation and batch normalization.Depthwise convolution uses one group per input channel, and ConvMixer features retain the same resolution and size across layers.
- Decoder stage with skip connection: The decoder uses five modules with convolution blocks and bilinear upsampling to reconstruct segmentation features.Each upsampling block includes an upsampling layer, convolution, batch normalization, and ReLU activation.
- Decoder stage with skip connection: Multi-scale attention gates suppress unimportant features and enhance valuable ones before combining encoder features with corresponding upsampled decoder features.The gates are integrated into skip connections for feature transfer and accurate positioning.
- Decoder stage with skip connection: Three convolutions—pointwise, ordinary 3×3, and dilated 3×3 with dilation rate 2—extract features at different receptive fields before concatenation.The resulting feature maps have the same size, enabling adaptive selection across resolutions.
3. RESULTS AND DISCUSSION
CMU-Net was evaluated on BUSI and TUS ultrasound datasets against six segmentation approaches using standard metrics and ablation studies. It achieved the strongest reported performance, with ConvMixer and multi-scale attention gates contributing to improvements.
- 3.1. Datasets, evaluation and implementation details: The loss combines binary cross entropy and Dice loss, and the experiments used Adam optimization with three random 80/20 train-validation splits.Images were resized to 256×256 and augmented with random rotation and flipping.
- 3.1. Datasets, evaluation and implementation details: CMU-Net was compared with U-Net, U-Net++, Attention U-Net, U-Net3+, TransUnet, and UNeXt on BUSI and TUS.The evaluation used the BUSI and private TUS datasets, with five quantitative metrics.
- 3.2. Results: CMU-Net outperformed six competing approaches, improving BUSI IoU by 2.89% and F1 score by 2.00%.On TUS, it improved all metrics and achieved a better trade-off between recall and specificity.
- 3.2. Results: The ConvMixer block with depth 7 and kernel size 7 achieved the best performance among the tested ConvMixer configurations.
- 3.3. Ablation study: Adding the ConvMixer block substantially improved the original U-Net, while multi-scale attention gates improved performance further.The attention gates magnified helpful encoder features during knowledge transfer.
4. CONCLUSION AND PERSPECTIVES
The paper concludes that CMU-Net is a fully convolutional ultrasound segmentation network combining ConvMixer global-context modeling with multi-scale attention gates. It reports state-of-the-art performance on two ultrasound datasets and identifies further directions for improving accuracy and interpretability.
- 4. CONCLUSION AND PERSPECTIVES: CMU-Net achieved state-of-the-art performance on two ultrasound datasets by combining ConvMixer blocks with multi-scale attention gates.
- 4. CONCLUSION AND PERSPECTIVES: Future work includes testing larger convolution kernels, placing ConvMixer blocks at different encoder levels, and analyzing errors.
- 4. CONCLUSION AND PERSPECTIVES: The authors propose incorporating lesion physiological and anatomical structures to improve model interpretability.
- 4. CONCLUSION AND PERSPECTIVES: Figure 4 presents BUSI and TUS segmentation predictions for CMU-Net and six comparison models alongside inputs and ground truth.