Source-linked AI summary
MDU-Net: Multi-scale Densely Connected U-Net for biomedical image segmentation
Jiawei Zhang, Yuzhen Jin, Jilan Xu, Xiaowei Xu, Yanchun Zhang
TL;DR
Biomedical image segmentation is difficult because target tissues vary histologically, while manual segmentation is time-consuming and expensive. The paper proposes multi-scale dense connections for U-shaped networks, forming MDU-Net with quantization, and reports improved segmentation performance over U-Net on the MICCAI 2015 GlaS dataset.
Problem
Biomedical image segmentation requires accurate pixel predictions despite diverse histological variation, and manual segmentation is time-consuming and expensive.
Method
The paper introduces multi-scale dense connections across U-shaped encoders, decoders, and cross-layer pathways, combining them in MDU-Net with quantization to reduce overfitting.
Results
The proposed dense structures improve U-Net accuracy by 2% on average on the MICCAI 2015 GlaS dataset, while the model achieves state-of-the-art accuracy with an ignorable parameter increment.
Takeaways & Limitations
MDU-Net provides an efficient U-shape-based backbone that strengthens feature aggregation and can extend beyond the original U-Net architecture.
Abstract
from arXiv · showhide
Biomedical image segmentation plays a central role in quantitative analysis, clinical diagnosis, and medical intervention. In the light of the fully convolutional networks (FCN) and U-Net, deep convolutional networks (DNNs) have made significant contributions to biomedical image segmentation applications. In this paper, we propose three different multi-scale dense connections (MDC) for the encoder, the decoder of U-shaped architectures, and across them. Based on three dense connections, we propose a multi-scale densely connected U-Net (MDU-Net) for biomedical image segmentation. MDU-Net directly fuses the neighboring feature maps with different scales from both higher layers and lower layers to strengthen feature propagation in the current layer. Multi-scale dense connections, which contain shorter connections between layers close to the input and output, also make a much deeper U-Net possible. Besides, we introduce quantization to alleviate the potential overfitting in dense connections, and further improve the segmentation performance. We evaluate our proposed model on the MICCAI 2015 Gland Segmentation (GlaS) dataset. The three MDC improve U-Net performance by up to 1.8% on test A and 3.5% on test B in the MICCAI Gland dataset. Meanwhile, the MDU-Net with quantization obviously improves the segmentation performance of original U-Net.
Introduction
Biomedical image segmentation is important for diagnosis, surgical planning, and treatment, but accurate pixel-level prediction is difficult because target tissues vary histologically and manual segmentation is costly. The paper introduces multi-scale dense connections that fuse feature maps from higher and lower layers to improve feature representation.
- Motivation: Biomedical segmentation supports diagnosis, surgical planning, and treatment, but target variation makes accurate high-level pixel prediction difficult.Traditional segmentation by experienced doctors is also time-consuming and expensive.
- Background: Deep convolutional networks built on FCN and U-Net have substantially improved biomedical image segmentation.
- Existing approaches: U-Net uses skip connections to reuse and fuse features, but prior dense-connection approaches have important fusion and computational limitations.U-Net++ fuses higher-resolution decoder features but incurs substantial computational cost and only combines features from higher layers.
- Proposed direction: Multi-scale dense connections resize learned features to a common resolution and fuse higher-layer semantic features with lower-layer fine-grained features.The operation uses 1 × 1 convolutions for channel adjustment and adds only small extra parameters.
Related work
Related work develops U-Net-based encoder-decoder architectures, dense connections, and multi-scale feature aggregation for segmentation and related dense-prediction tasks. Existing methods use strategies such as stacked U-Nets, residual structures, pyramids, atrous convolutions, deformable convolutions, and adaptive preprocessing, motivating the paper’s multi-scale connectivity pattern.
- U-Net architectures: U-Net-based models commonly use encoder-decoder architectures to recover high-resolution representations for segmentation and other tasks.Stacked U-Nets iteratively fuse multi-scale features without changing their resolutions.
- Network design: Width, depth, residual shortcuts, and dense connections have been explored to improve feature fusion, reuse, optimization, and multi-scale processing.Examples include Inception modules, residual networks, and denser decoder or scale-sampling structures.
- Multi-scale aggregation: Multi-scale context methods include image pyramids, dilated convolutions, ASPP, Dense-ASPP, and deformable convolutions.These methods target features or receptive fields spanning different object scales.
- Biomedical applications: Biomedical image segmentation has shifted from hand-crafted and graph-based methods toward deep learning because malignant subjects vary substantially in appearance.Recent systems address varying boundaries, sizes, semantic correlations, and dataset-specific hyperparameters.
Method
MDU-Net extends U-Net with multi-scale dense connections in the encoder, decoder, and across them. These connections fuse resized features across resolutions, while the architecture also includes fully dense variants and quantization strategies.
- Multi-scale dense connections: MDU-Net combines multi-scale dense connections in the encoder, decoder, and across the U-shaped network.The framework is based on U-Net and incorporates three dense-connection types.
- Encoder and decoder blocks: Each convolutional block uses two cascaded 3 × 3 convolutions followed by batch normalization and ReLU, with the decoder block constructed similarly.The example dense encoder unit uses n = 2.
- Configurations and efficiency: The method evaluates alternative encoder, decoder, cross-connection, and quantization configurations within the MDU-Net family.The supplied tables compare these architectural and quantization variants, including parameter-efficiency comparisons.
- Cross connections: Cross connections link encoder and decoder features with different resolutions, including architectures that selectively fuse higher- or lower-resolution encoder features.The cross-connected features combine information from the encoder and decoder pathways.
- Fully dense architecture: The fully dense U-shaped architecture combines multi-scale dense cross connections with a multi-scale dense decoder and re-encodes the resulting information.The resulting feature maps share the same number of channels as the original features.
Experiments
Experiments evaluate MDU-Net and its dense-connection variants on the MICCAI 2015 GlaS dataset, examining connection structure, efficiency, and quantization. Increasing dense connections generally improves segmentation, while partial quantization mitigates overfitting without the accuracy loss of full quantization.
- Experiment setup: The evaluation uses 165 GlaS histology images, with 85 for training and 80 for testing across Test Parts A and B.The experiments also vary dense-connection configurations and assess quantization effects.
- Dense-connection analysis: Increasing the number of dense connections generally improves accuracy, with modified structures achieving a 2% average superiority over U-Net on GlaS.Decoder dense connections improve slightly more than encoder and cross connections.
- Dense-connection analysis: The experiments combine encoder, decoder, and cross dense blocks to identify an effective MDU-Net configuration.Encoder4, decoder4, and cross5 are selected for combination experiments.
- Network efficiency: MDU-Net achieves state-of-the-art accuracy with an ignorable parameter increase over U-Net.The dense structure adds parameters primarily through 1 × 1 convolutions used for feature reuse and concatenation.
- Quantization: Half quantization reduces overfitting, whereas full quantization decreases segmentation performance; the final model therefore uses half-quantized weights.The half-quantized model reaches 0.88 accuracy on Test B.
Conclusion
The paper proposes MDU-Net by combining multi-scale dense connections across the encoder, decoder, and cross-layer pathways, with quantization used to reduce overfitting. On GlaS, the model improves the Dice coefficient over U-Net by up to 3% on Test A and 4.1% on Test B.
- Conclusion: MDU-Net combines three multi-scale dense architectures with quantization to strengthen feature aggregation and reduce overfitting.The connections fuse higher-layer and lower-layer feature maps at the current layer.
- Conclusion: The Dice coefficient improves over U-Net by up to 3% on Test A and 4.1% on Test B.