Source-linked AI summary
Using DUCK-Net for Polyp Image Segmentation
Razvan-Gabriel Dumitru, Darius Peteleaza, Catalin Craciun
TL;DR
Polyp segmentation is clinically important but difficult, while deep-learning systems commonly depend on substantial labeled data. The paper introduces DUCK-Net, a supervised encoder-decoder CNN with custom convolutions, residual downsampling, and augmentation, and reports strong benchmark performance and generalization from limited training data. Its main scope boundary is reduced accuracy for polyps whose colors blend into the background and obscure their borders.
Problem
Polyp segmentation is challenging because polyps vary in shape, size, and texture, while deep-learning methods typically require costly and time-consuming labeled datasets.
Method
DUCK-Net combines a U-Net-style encoder-decoder, custom DUCK convolutional blocks, residual downsampling, and data augmentation for polyp segmentation.
Results
DUCK-Net achieves state-of-the-art performance across multiple polyp-segmentation benchmarks and demonstrates strong generalization with limited training data.
Takeaways & Limitations
The architecture shows potential for polyp segmentation and other image-segmentation tasks when training data are limited.
Takeaways & Limitations
Performance is limited for polyps whose colors blend with the background, producing indistinct borders that require further investigation.
Abstract
from arXiv · showhide
This paper presents a novel supervised convolutional neural network architecture, "DUCK-Net", capable of effectively learning and generalizing from small amounts of medical images to perform accurate segmentation tasks. Our model utilizes an encoder-decoder structure with a residual downsampling mechanism and a custom convolutional block to capture and process image information at multiple resolutions in the encoder segment. We employ data augmentation techniques to enrich the training set, thus increasing our model's performance. While our architecture is versatile and applicable to various segmentation tasks, in this study, we demonstrate its capabilities specifically for polyp segmentation in colonoscopy images. We evaluate the performance of our method on several popular benchmark datasets for polyp segmentation, Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and ETIS-LARIBPOLYPDB showing that it achieves state-of-the-art results in terms of mean Dice coefficient, Jaccard index, Precision, Recall, and Accuracy. Our approach demonstrates strong generalization capabilities, achieving excellent performance even with limited training data. The code is publicly available on GitHub: https://github.com/RazvanDu/DUCK-Net
Introduction
Polyp segmentation is clinically important but difficult because polyps vary widely and labeled medical images are costly to obtain. DUCK-Net addresses these challenges with a supervised U-Net-based architecture combining custom convolutions and residual downsampling.
- Clinical motivation: Colorectal cancer is a leading cause of global cancer mortality, and early polyp detection and removal support prevention.Colonoscopy is the gold standard, but missed or misdiagnosed polyps can seriously harm patients.
- Clinical motivation: Polyp segmentation is difficult because colonoscopy images contain polyps with diverse shapes, sizes, and textures.This diversity can lead to missed or misdiagnosed polyps.
- Research gap: Deep learning methods can achieve high precision but typically require large, costly, and time-consuming labeled datasets.Existing architectures may also be less robust across varied polyp characteristics.
- Proposed approach: DUCK-Net uses a U-Net encoder-decoder with a custom convolutional block and residual downsampling to improve multiresolution polyp-feature processing.The architecture is designed to locate polyp borders accurately while retaining initial image information at each encoder resolution.
- Proposed approach: The method combines DeepLabV3 atrous convolutions and a ResUNet++ residual block for spatial-information capture and feature extraction.Its stated contributions also include training without external modules or pre-training and evaluation across varying polyp appearances.
Related work
Prior polyp-segmentation research spans CNN, attention, multiscale, and transformer-based architectures, each balancing accuracy, robustness, computation, and data requirements. DUCK-Net is positioned within this active field as a traditional CNN approach exploring improvements without relying on ViT-based encoders.
- CNN architectures: U-Net is a simple, efficient biomedical encoder-decoder, but it may struggle with complex or highly varied input images.This motivates alternative architectures for challenging segmentation settings.
- CNN architectures: PraNet combines global guidance, reverse attention, and recurrent cooperation to improve polyp-boundary modeling and reported segmentation efficiency.Its evaluations reported approximately 50 frames per second.
- Multiscale and attention methods: DeepLabV3+ uses atrous convolutions to capture multiscale and long-range context for complex shapes, at the cost of greater computation and slower training or inference.The architecture extends DeepLabV3 for semantic image segmentation.
- Multiscale and attention methods: HRNetV2 shares information across resolutions and can improve performance on small or blurry objects, but may overfit and require more data.ResUNet and HarDNet-DFUS respectively emphasize residual feature extraction and real-time prediction, with different accuracy, data, and computation trade-offs.
- Multiscale and attention methods: ColonFormer uses encoder and refinement attention mechanisms, which can handle complex inputs but may require more computation and be harder to optimize.MSRF-Net instead exchanges multiscale features through Dual-Scale Dense Fusion blocks to preserve resolution and information flow.
- Transformer-based methods: Transformer-based methods commonly use pretrained vision-transformer encoders and have achieved state-of-the-art results on Kvasir-SEG.Their use of global context motivates comparison with traditional CNN approaches.
- Research direction: Traditional CNNs can remain competitive with vision transformers, motivating DUCK-Net’s exploration of CNN-based improvements for biomedical segmentation.The field still requires careful evaluation of accuracy, computational efficiency, and training strategies.
Methodology
DUCK-Net combines a U-Net encoder-decoder with a custom parallel DUCK block and residual downsampling to capture features at multiple scales while retaining low-level information.
- Architecture: A secondary downscaling path preserves fine details by bypassing convolutional processing and adding its outputs to the main downsampling path.The path uses learned 2D 2x2 convolutions with stride 2 instead of max pooling.
- Architecture: The architecture combines decoder outputs by addition rather than concatenation to reduce memory and computational requirements.The authors report that addition produced better results and requires fewer upsampling parameters.
- DUCK components: Residual blocks capture small polyp details, while Midscope and Widescope blocks simulate 7x7 and 15x15 kernels with dilated convolutions.Residual-block combinations simulate 5x5, 9x9, and 13x13 kernels.
- DUCK components: The Separated block approximates NxN kernels through sequential 1xN and Nx1 convolutions but may lose diagonal spatial relationships.The other DUCK components are designed to compensate for this loss of diagonality.
- Architecture: DUCK-Net replaces conventional U-Net convolutional pairs with a custom block that simulates varied kernel sizes in parallel.The block combines residual, dilated, and separated convolutional designs to capture low- and high-level features.
- Training and evaluation: Performance is evaluated with Dice Coefficient, Jaccard Index, Precision, Recall, and Accuracy, using Dice loss for segmentation training.The paper describes Dice loss as suitable for overlap optimization and class-imbalance handling.
Experiments
The experiments evaluate DUCK-Net on four polyp-segmentation datasets using controlled splits, augmentation, cross-dataset testing, and ablation studies. DUCK consistently outperformed a simple convolutional block across the tested metrics on Kvasir-SEG.
- Experimental setup: All models used identical random 80:10:10 training, validation, and testing splits to support fair comparisons.The datasets and splits were provided for reproducibility.
- Generalization tests: The study tested both within-dataset performance and generalization by training on Kvasir-SEG and testing on CVC-ClinicDB, and vice versa.The setup was designed to assess performance on unseen data and across contexts.
- Training procedure: Data augmentation was applied before each epoch, with geometric transformations shared by images and masks while color jitter affected images only.The augmentation pipeline included flips, color jitter, affine transformations, rotations, translations, scaling, and shearing.
- Datasets and metrics: The evaluation covered Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and ETIS-LARIBPOLYPDB, using Dice, Jaccard, Precision, Recall, Accuracy, and standard deviation.The standard deviation was included to assess metric variability across models.
- Ablation study: The DUCK block consistently outperformed a simple convolutional block on Kvasir-SEG in both the 17- and 34-filter models across all tested metrics.The controlled comparison supports integrating the DUCK block into DUCK-Net.
Discussion
The discussion reports strong benchmark performance and generalization for DUCK-Net, while identifying computational-resource demands and difficulty with low-contrast polyp borders. The DUCK block’s ablation results further support its contribution to segmentation performance.
- Architecture: DUCK-Net combines wide information extraction from atrous convolutions with rich information extraction from a large efficient separable kernel to localize polyps.The discussion connects these design choices with accurate target localization.
- Benchmark performance: DUCK-Net outperformed other architectures on four benchmark datasets and learned key polyp features from small amounts of data.The benchmarks were Kvasir-SEG, CVC-ClinicDB, ETIS-LaribPolypDB, and CVC-ColonDB.
- Generalization: The model generalized between Kvasir-SEG and CVC-ClinicDB, although FCN-Transformer achieved stronger generalization with additional pre-training data.The comparison concerns cross-dataset results rather than within-dataset benchmark performance.
- Ablation evidence: The DUCK block consistently outperformed a simple convolutional block in ablation studies, supporting its effectiveness within the DUCK-Net architecture.The paper suggests evaluating the block in other architectures and tasks in future work.
- Limitations: Additional convolutional blocks increase computational cost and memory usage, constraining deployment in resource-limited environments.Dilated and separable convolutions are discussed as lower-cost alternatives with similar representational power.
- Limitations: Performance is limited for polyps whose colors blend into the background and produce indistinct borders.The paper calls for further investigation to improve border localization in these cases.
Conclusion
DUCK-Net achieves state-of-the-art performance for polyp segmentation in colonoscopy images while processing image information at multiple resolutions. It also demonstrates strong generalization and excellent results with limited training data.
- The encoder-decoder architecture combines residual downsampling and a custom convolutional block to process image information at multiple resolutions effectively.
- Data augmentation techniques help improve DUCK-Net’s overall performance.
- DUCK-Net demonstrates strong generalization capabilities and achieves excellent results even with limited training data.