Source-linked AI summary
DC-UNet: Rethinking the U-Net Architecture with Dual Channel Efficient CNN for Medical Images Segmentation
Ange Lou, Shuyue Guan, Murray Loew
TL;DR
Classical U-Net has limitations for challenging medical-image segmentation, motivating an enhanced architecture. DC-UNet replaces the encoder-decoder with an efficient CNN and the skip connection with a residual module, achieving reported improvements over U-Net on three datasets.
Problem
Classical U-Net has limitations in challenging medical-image segmentation tasks, including cases with vague object boundaries.
Method
DC-UNet uses an efficient CNN in place of the encoder and decoder and a residual module in place of the encoder-decoder skip connection.
Results
DC-UNet achieved relative segmentation-accuracy improvements of 2.90%, 1.49%, and 11.42% over U-Net on three challenging datasets.
Takeaways & Limitations
DC-UNet also improved over Multi-ResUNet by 1.20%, 0.66%, and 2.02%, while being described as more reliable and robust.
Takeaways & Limitations
The paper states that DC-UNet still has limitations related to classical U-Net.
Abstract
from arXiv · showhide
Recently, deep learning has become much more popular in computer vision area. The Convolution Neural Network (CNN) has brought a breakthrough in images segmentation areas, especially, for medical images. In this regard, U-Net is the predominant approach to medical image segmentation task. The U-Net not only performs well in segmenting multimodal medical images generally, but also in some tough cases of them. However, we found that the classical U-Net architecture has limitation in several aspects. Therefore, we applied modifications: 1) designed efficient CNN architecture to replace encoder and decoder, 2) applied residual module to replace skip connection between encoder and decoder to improve based on the-state-of-the-art U-Net model. Following these modifications, we designed a novel architecture--DC-UNet, as a potential successor to the U-Net architecture. We created a new effective CNN architecture and build the DC-UNet based on this CNN. We have evaluated our model on three datasets with tough cases and have obtained a relative improvement in performance of 2.90%, 1.49% and 11.42% respectively compared with classical U-Net. In addition, we used the Tanimoto similarity to replace the Jaccard similarity for gray-to-gray image comparisons.
I. INTRODUCTION
Medical image segmentation is important but challenging cases expose limitations in traditional methods and classical U-Net. The paper introduces DC-UNet and reports improved performance across challenging medical-image datasets.
- Medical images contain substantial data and noise interference, making effective image processing and information extraction important.
- Traditional segmentation methods do not suit challenging tasks such as polyps with highly variable shapes, sizes, and vague boundaries.
- Deep learning segmentation methods include region-based approaches and fully convolutional networks, with U-Net among the encoder-decoder models used for biomedical images.
- DC-UNet is introduced as an enhanced U-Net version for medical image segmentation.
- DC-UNet surpasses classical U-Net in all tested cases while using slightly fewer parameters.
II. METHODS
The method retains U-Net’s encoder-decoder segmentation structure while replacing its standard components with multiscale convolutional blocks and residual paths. These changes target multiscale feature extraction, computational efficiency, and spatial-feature transfer.
- U-Net architecture: U-Net uses an encoder to extract spatial features and a decoder to reconstruct the segmentation map.The encoder repeatedly applies convolution and pooling, while the decoder upsamples and reduces channels before producing the final map.
- U-Net architecture: U-Net skip connections concatenate encoder outputs with upsampled decoder features to retrieve spatial information lost during pooling.
- MultiResUNet architecture: Inception-like blocks apply parallel convolutions with varying kernel sizes to extract features at different scales.1 × 1 convolutions reduce dimensions before the 3 × 3 and 5 × 5 convolutions.
- MultiResUNet architecture: Replacing larger convolutions with successive 3 × 3 layers reduces computational cost while preserving a comparable receptive field.Three 3 × 3 convolutions provide the same receptive field as a 7 × 7 convolution with lower stated cost.
- MultiResUNet architecture: The MultiRes block combines multiscale convolutional features with a residual connection, giving U-Net multiresolution analysis and additional spatial features.
- MultiResUNet architecture: Res-Path replaces the encoder-decoder skip connection with chained 3 × 3 convolutions and residual connections before concatenating decoder features.
C. DC-UNet
DC-UNet addresses challenging medical-image segmentation cases by introducing a Dual-Channel block and combining it with Res-Path connections. The design targets insufficient spatial features and interference from fuzzy objects or backgrounds.
- Comparison: The section compares classical U-Net, MultiResUNet, and DC-UNet while describing MultiResUNet as providing better output than U-Net.The comparisons are framed around segmentation performance and parameter advantages.
- Motivation and high-level considerations: MultiResUNet can struggle with extremely challenging cases involving fuzzy objects and background interference from medical equipment.These factors can influence segmentation results, including in breast infrared images with unclear boundaries.
- Dual-Channel block: The authors modified the MultiRes block to provide more effective features for separating objects from whole images.This modification led to the construction of a new block for improvement.
- Motivation and high-level considerations: Different-scale features were observed to improve segmentation, motivating efforts to provide more effective spatial features.The authors connect this motivation to limitations of the residual connection in the MultiRes block.
- Dual-Channel block: A sequence of three 3 × 3 convolutional layers replaces the MultiRes residual connection in the proposed Dual-Channel block.The Dual-Channel block is illustrated in Fig. 7.
- DC-UNet architecture: DC-UNet combines the Res-Path connection between encoder and decoder with the Dual-Channel block.The resulting architecture is presented as a new U-Net architecture, with its architecture illustrated in Fig. 8.
III. EXPERIMENTS
The experiments implemented the network models in Python using Keras with a TensorFlow backend. They were conducted on a desktop computer equipped with an Intel Core i7-9700K CPU, 16.0 GB RAM, and an NVIDIA GeForce RTX 2070 GPU.
- Experimental setup: The network models were built with Keras and a TensorFlow backend in Python 3.The implementation used Keras, TensorFlow, and Python 3.
- Experimental setup: The experiments ran on a desktop computer with an Intel Core i7-9700K processor and 16.0 GB of RAM.The processor operates at 3.6 GHz.
- Experimental setup: The desktop computer used an NVIDIA GeForce RTX 2070 GPU for the experiments.
A. Baseline model
The experiments compare U-Net, MultiResUNet, and DC-UNet under matched five-stage encoder–decoder settings. They also describe image preprocessing, binary cross-entropy training, and Adam optimization.
- Model comparison: U-Net was selected as the baseline and compared with MultiResUNet and DC-UNet.The stated experimental goal was to evaluate DC-UNet against both reference architectures.
- Model configurations: The classical U-Net used five encoder and decoder stages with filter numbers {64, 128, 256, 512, 1024}.
- Model configurations: MultiResUNet and DC-UNet also used five encoder and decoder stages, with layer filter numbers specified in Tables 1 and 2.
- Preprocessing: Thermography images were converted from 16-bit to 8-bit and resized to 256 × 128 before processing.Other databases were resized so image width and height were no larger than 256 because of GPU-memory limits.
- Training objective: Because semantic segmentation is treated as pixel-wise binary classification, the models used binary cross-entropy as the loss function.The passage defines X as the input image, ŷ as the model prediction, and y as the ground truth.
- Optimization: The models were trained with the Adam optimizer using β1 = 0.9 and β2 = 0.999.The supplied passage begins the training-configuration description with these Adam parameters.
D. Measurement metric
The paper compares binary and grayscale image-similarity measures for segmentation evaluation, selecting Tanimoto similarity for grayscale outputs. Experiments indicate that Tanimoto remains stable under image-size and object-ratio changes while producing values close to Jaccard similarity.
- Tanimoto similarity extends Jaccard similarity to grayscale image comparisons, whereas Jaccard similarity is intended for binary images.The paper uses Tanimoto similarity as an alternative to Jaccard similarity for grayscale segmentation outputs.
- The evaluation compares Jaccard similarity, MAE, Tanimoto similarity, and SSIM across image sizes and object-area ratios.Images are resized by down-sampling, while blank margins alter the object-area ratio; results average 15 samples from one patient.
- For 9/15 samples at 200x400, Tanimoto similarity values are close to Jaccard similarity values.This comparison is reported for the 15 samples evaluated in Figure 12.
- Binarizing grayscale segmentation outputs can lose information, motivating direct grayscale-to-grayscale comparison.The paper notes that neural-network segmentation outputs are grayscale and that binarization may discard information.
E. Cross-validation
The study uses k-fold cross-validation to estimate model performance across repeated training and validation splits. It evaluates three U-Net-based models on two public datasets and an infrared breast dataset, which contains challenging acquisition conditions and background interference.
- In k-fold cross-validation, the dataset is divided into k mutually exclusive subsets of approximately equal size.The model runs k times, using one subset for validation and the remaining subsets for training each time.
- Overall model performance is estimated from the results obtained across all k training runs.
- The experiments use two public datasets and the authors’ infrared breast dataset to test three U-Net-based models.
- The infrared breast dataset contains 450 images from 14 patients and 16 healthy volunteers, with background objects and noise in every image.
- Infrared images were resized to 256 × 128 because of memory limitations.
B. Electron microscopy (EM)
The paper evaluates DC-UNet on medical-image datasets, including infrared breast images and endoscopy images, using resized inputs and cross-validation experiments. Reported examples show higher segmentation accuracy and clearer separation of structures in difficult cases.
- The study evaluates DC-UNet on three medical datasets and reports model parameter numbers for the experiments.
- The infrared breast dataset contains 450 images and is divided into 30 subsets by participant for cross-validation.
- After 30-fold cross-validation, average accuracies are 89.80% for U-Net, 91.47% for MultiResUNet, and 92.71% for DC-UNet.
- In a simple breast-segmentation case, accuracies are 92.47% for U-Net, 93.86% for MultiResUNet, and 95.38% for DC-UNet.
- In a challenging patient case with medical-equipment and body interference, accuracies are 86.47% for U-Net, 84.01% for MultiResUNet, and 92.62% for DC-UNet.
C. Results of electron microscopy image
The reported cross-validation results show DC-UNet outperforming the comparison models on EM and endoscopy segmentation cases. Visual examples associate its results with clearer boundaries, recovered separating lines, and improved handling of vague boundaries and small objects.
- Electron microscopy (EM): DC-UNet gives the best results for all reported EM cases.The EM comparison uses 5-fold cross-validation and records Tanimoto accuracy after 50 training epochs per run.
- Electron microscopy (EM): On EM images, DC-UNet captures separating lines missed by U-Net and MultiResUNet and produces clearer segmentation results.
- Endoscopy: On the endoscopy dataset, DC-UNet improves average accuracy by 9.4% over U-Net.
- Endoscopy: DC-UNet successfully segments images with vague boundaries and detects small objects in the endoscopy examples.
- Endoscopy: Endoscopy accuracy improves by 11.42% relative to U-Net, while MultiResUNet performs poorly on some challenging tasks.
- Endoscopy: For some easy endoscopy cases, MultiResUNet and DC-UNet perform similarly and both outperform classical U-Net.
VI. DISCUSSION
DC-UNet shows potential advantages on challenging medical image segmentation cases, combining stronger segmentation behavior with fewer parameters than U-Net and MultiResUNet. The discussion also identifies data augmentation, image enhancement, and broader evaluation as future directions.
- DC-UNet separates regions with similar temperatures and provides more accurate contours on the infrared breast dataset.
- DC-UNet produces less noisy results than MultiResUNet and U-Net on the EM dataset.
- DC-UNet shows strong ability on CVC-ClinicDB, where polyp boundaries, shapes, sizes, structures, and positions are challenging.
- DC-UNet uses fewer parameters because its dual-channel design reduces intermediate dimensions while retaining doubled multi-resolution features.Each channel uses half the corresponding MultiRes-block filter number, and channels are summed rather than concatenated.
- Data augmentation can help avoid overfitting, while histogram equalization and CLAHE can improve contrast for unclear boundaries caused by illumination, noise, and tissue properties.
- Future work will test DC-UNet on more datasets and study how data augmentation and preprocessing affect performance.
VII. CONCLUSION
The paper develops DC-UNet as a U-Net-like architecture using a dual-channel CNN and evaluates it on three challenging medical image datasets. It reports higher segmentation accuracy and more reliable handling of vague boundaries, noise, and fine details than the compared models, while acknowledging remaining limitations.
- Visual results were closer to ground truth, while U-Net and MultiResUNet sometimes under-segmented or completely missed objects.
- The authors designed DC-UNet after analyzing classical U-Net and MultiResUNet, aiming to provide more effective features with fewer parameters.
- The evaluation uses two public medical datasets and an infrared breast dataset containing challenging cases such as unclear boundaries, noise, and difficult polyp shapes.
- DC-UNet achieved 1.20%, 0.66%, and 2.02% improvements over MultiResUNet across the evaluated datasets.
- DC-UNet detected vague boundaries, reduced noise interference, and captured fine details in challenging cases.
- The conclusion presents DC-UNet as an effective medical image segmentation model but notes that classical U-Net and MultiResUNet still have limitations.