Source-linked AI summary
DDANet: Dual Decoder Attention Network for Automatic Polyp Segmentation
Nikhil Kumar Tomar, Debesh Jha, Sharib Ali, Håvard D. Johansen, Dag Johansen, Michael A. Riegler, Pål Halvorsen
TL;DR
Polyp segmentation can provide lesion-boundary information for clinical analysis, but manual assessment takes expert time and may miss lesions. DDANet uses a shared encoder with segmentation and autoencoder decoders to segment polyps. On the unseen test dataset, it achieved a DSC of 0.7874 and averaged 70.23 FPS.
Problem
Manual polyp assessment takes expert time and can involve missed lesions, while segmentation can provide detailed boundaries for clinical analysis.
Source: §1 · p. 1Method
DDANet uses a shared encoder with parallel segmentation and autoencoder decoders, with squeeze-and-excitation attention to reweight feature channels.
Source: §1 · p. 2, §3.3 · p. 4, §3.2 · p. 4Results
On the unseen test dataset, DDANet achieved a DSC of 0.7874 and averaged 70.23 FPS.
Source: §6 · p. 7Takeaways & Limitations
The authors report that DDANet identifies polyps in real time at an average of 70.23 FPS.
Source: §6 · p. 7Qualitative results still show challenges with flat or sessile polyps.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Colonoscopy is the gold standard for examination and detection of colorectal polyps. Localization and delineation of polyps can play a vital role in treatment (e.g., surgical planning) and prognostic decision making. Polyp segmentation can provide detailed boundary information for clinical analysis. Convolutional neural networks have improved the performance in colonoscopy. However, polyps usually possess various challenges, such as intra-and inter-class variation and noise. While manual labeling for polyp assessment requires time from experts and is prone to human error (e.g., missed lesions), an automated, accurate, and fast segmentation can improve the quality of delineated lesion boundaries and reduce missed rate. The Endotect challenge provides an opportunity to benchmark computer vision methods by training on the publicly available Hyperkvasir and testing on a separate unseen dataset. In this paper, we propose a novel architecture called ``DDANet'' based on a dual decoder attention network. Our experiments demonstrate that the model trained on the Kvasir-SEG dataset and tested on an unseen dataset achieves a dice coefficient of 0.7874, mIoU of 0.7010, recall of 0.7987, and a precision of 0.8577, demonstrating the generalization ability of our model.
1 Introduction
Polyp segmentation can provide clinically important lesion-boundary information, while manual assessment takes expert time and may miss lesions. DDANet is proposed for the Endotect challenge’s automatic segmentation task on an unseen dataset.
1 Introduction
- Adenoma miss rates of 6 to 27% and the clinical relevance of identifying lesion pixels motivate automatic polyp segmentation.The introduction links colonoscopy screening and lesion removal with early cancer detection and prevention.Source: §1 · p. 1
- Automated polyp detection or surveillance could provide low-cost solutions and save clinicians time for more severe cases.Source: §1 · p. 2
- The Endotect segmentation task asks participants to generate automatic polyp masks for an unseen dataset.Source: §1 · p. 2
- DDANet uses a shared encoder with parallel segmentation and autoencoder decoders, whose attention maps strengthen encoder features.The network predicts both a segmentation mask and a reconstructed grayscale image.Source: §1 · p. 2
2 Related Work
The paper situates polyp segmentation in a field where diverse datasets and hardware make comparisons difficult. It presents challenge benchmarking and an autoencoder-augmented architecture as relevant approaches.
2 Related Work
- Different datasets and hardware make comparisons among polyp-segmentation approaches difficult, while competitions benchmark methods on common data.Challenge organizers also calculate evaluation metrics independently, including team rankings.Source: §2 · p. 2
- Competitions can help identify methods’ strengths and weaknesses and provide a shared space to discuss results.Source: §2 · p. 2
- DDANet adds an autoencoder branch to encoder-decoder networks such as UNet, ResUNet++ [13], and DoubleUNet.The paper states that the branch’s benefit is reflected in quantitative and qualitative results.Source: §2 · p. 3
3 DDANet
DDANet combines residual learning and channelwise squeeze-and-excitation attention in a fully convolutional design with one encoder shared by two decoders.
3.1 Residual block
- Residual blocks use two 3×3 convolutions with batch normalization and ReLU, plus an identity shortcut that supports gradient flow.Source: §3.1 · p. 4
3.2 Squeeze and Excitation block
- Squeeze-and-excitation attention compresses feature maps by global average pooling, then uses a two-layer network to scale feature channels.Source: §3.2 · p. 4
3.3 The DDANet architecture
- DDANet combines residual learning and squeeze-and-excitation in a fully convolutional network with one encoder shared by two four-block decoders.Source: §3.3 · p. 4
- Both decoders receive encoder features through upsampling and skip connections; the autoencoder branch generates attention that modulates segmentation features.The segmentation decoder outputs a mask, while the autoencoder decoder outputs a reconstructed grayscale image.Source: §3.3 · pp. 4–5
4 Experimental Setup
The experimental setup reports implementation choices and the Kvasir-SEG training and test data.
4.1 Implementation Details
- The implementation uses 512 × 512 inputs, 200 training epochs, and Adam with a learning rate of 1e−4.Training used PyTorch 1.6 on an NVIDIA DGX-2 with V100 Tensor Core GPUs; mask loss combined binary cross-entropy and dice loss.Source: §4.1 · p. 5
4.2 Datasets
- Kvasir-SEG contains 1,000 polyp images, and 88% were used for training while 12% formed the development-test set.Source: §4.2 · p. 5
- A separate organizer-provided test set contained 200 images, but its ground-truth masks were not provided.Source: §4.2 · p. 5
5 Results
DDANet reports higher DSC than prior results on Kvasir-SEG, though differing train-test splits prevent direct comparison. Qualitative outputs include large and small polyps, while flat polyps remain challenging.
- On Kvasir-SEG, DDANet obtained a DSC of 0.8576, mIoU of 0.7800, recall of 0.8880, and precision of 0.8643.Source: §5 · p. 5, Table 1 · p. 6
- Figure 2 includes a reconstruction component used as an auxiliary task alongside segmentation.Source: Figure 2 · p. 6
- 0.8576 DSC exceeds prior reported values 0.8133 and 0.7877, but differing train-test splits prevent direct comparison.Source: §5 · p. 6
- The qualitative results show segmentation of both large and small polyps, but flat polyps remain challenging.Source: §5 · p. 6
- The authors describe segmentation outputs as reliable across development and unseen test results.Source: §5 · p. 7
6 Discussion
On the unseen dataset, DDANet achieved a DSC of 0.7874 and an average 70.23 FPS; qualitative results show remaining challenges with flat or sessile polyps.
- Qualitative results show segmentation across polyp sizes, but challenges remain for flat or sessile polyps.Source: §6 · p. 7
- The authors propose image super-resolution as future work in place of grayscale image reconstruction.Source: §6 · p. 7
- 0.7874 DSC on the unseen dataset and an average 70.23 FPS accompany the authors’ claim of real-time polyp identification.Source: §6 · p. 7
7 Conclusion
The authors report strong segmentation-task results for DDANet in the Endotect challenge and identify post-processing and parameter analysis as future improvement directions.