Source-linked AI summary
UACANet: Uncertainty Augmented Context Attention for Polyp Segmentation
Taehun Kim, Hyemin Lee, Daijin Kim
TL;DR
Polyp segmentation needs precise delineation of potentially cancerous tissue, but boundary information is difficult to obtain without costly edge annotations. UACANet adds uncertain saliency regions to context attention within a modified U-Net architecture and evaluates the approach across five benchmarks. It achieves state-of-the-art performance, including 76.6% mean Dice on ETIS, a 13.8% improvement over the previous state-of-the-art method.
Problem
Polyp segmentation supports early diagnosis of potentially cancerous polyps, while explicit edge guidance for boundary information can require expensive additional annotations.
Method
UACANet uses a modified U-Net with additional encoder-decoder paths and computes foreground, background, and uncertain saliency maps for context attention.
Results
UACANet achieves state-of-the-art performance across five polyp benchmarks, including 76.6% mean Dice on ETIS, 13.8% above the previous state-of-the-art method.
Takeaways & Limitations
Uncertain saliency regions can represent complementary boundary information without extra boundary guidance in the evaluated polyp segmentation benchmarks.
Takeaways & Limitations
The method assumes that boundary regions correspond to saliency scores around 0.5 and that saliency and reverse saliency contain similar edge information.
Abstract
from arXiv · showhide
We propose Uncertainty Augmented Context Attention network (UACANet) for polyp segmentation which consider a uncertain area of the saliency map. We construct a modified version of U-Net shape network with additional encoder and decoder and compute a saliency map in each bottom-up stream prediction module and propagate to the next prediction module. In each prediction module, previously predicted saliency map is utilized to compute foreground, background and uncertain area map and we aggregate the feature map with three area maps for each representation. Then we compute the relation between each representation and each pixel in the feature map. We conduct experiments on five popular polyp segmentation benchmarks, Kvasir, CVC-ClinicDB, ETIS, CVC-ColonDB and CVC-300, and achieve state-of-the-art performance. Especially, we achieve 76.6% mean Dice on ETIS dataset which is 13.8% improvement compared to the previous state-of-the-art method. Source code is publicly available at https://github.com/plemeri/UACANet
1 INTRODUCTION
Polyp segmentation supports early diagnosis because polyps can be potentially cancerous, while existing approaches borrow salient-object methods and often rely on costly edge guidance. UACANet addresses this by incorporating uncertain saliency regions with foreground and background context attention, achieving state-of-the-art results across five benchmarks.
- Motivation: Polyp detection in colonoscopy images can aid early diagnosis because polyps are potentially cancerous and require long-term monitoring.Monitoring includes polyp growth, location, and possible malignancy.
- Related work: Medical polyp segmentation methods commonly adapt salient-object detection because both tasks emphasize salient regions over surrounding scenes.Salient-object methods focus on classifying pixels belonging to the most salient object.
- Related work: Current salient-object methods often use edge guidance, but obtaining additional edge annotations can be expensive.Reverse attention provides boundary cues from saliency maps, although performance with and without it was reported as similar.
- Proposed approach: UACANet augments saliency-based context attention with foreground, background, and uncertain regions, using a modified U-Net with additional encoders and a decoder.The uncertain region is associated with ambiguous saliency scores and boundary information.
- Results: UACANet achieves state-of-the-art performance on Kvasir, CVC-ClinicDB, ETIS, CVC-ColonDB, and CVC-300.The supplied introduction states the cross-benchmark result but does not provide individual scores.
2 RELATED WORK
Related work spans semantic and salient-object segmentation, polyp-specific architectures, and boundary-aware refinement. These studies motivate UACANet’s use of multi-scale features, additional paths, and uncertain saliency context without relying on explicit edge annotations.
- Semantic segmentation: Semantic segmentation architectures use fully convolutional prediction, multi-scale features, pooling, and upsampling to classify pixels while recovering spatial information.FCN, deconvolution networks, PSPNet, and related architectures provide these design foundations.
- Salient object detection: Salient-object detection identifies important objects rather than predefined semantic classes, making accurate object prioritization difficult.Its criterion differs from semantic segmentation because it focuses on the object itself and surrounding contrast.
- Salient object detection: Boundary-aware salient-object methods use edge branches or reverse attention to recover complementary details, but edge detection may include redundant unrelated edges.Reverse attention uses the complement of a prediction to capture residual details for saliency refinement.
- Polyp segmentation: Polyp segmentation requires precise spatial detail in colonoscopy images, while multi-scale features placed late in the backbone may inadequately recover input details.DeeplabV3+ addresses this partly by concatenating low-level backbone features.
- Polyp segmentation: Polyp-specific networks include ResUNet++, SFA, and PraNet, which add specialized modules or paths for residual, boundary, and low-level feature handling.PraNet combines low-level features because its predecessor lacked sufficient sharing between top-down and bottom-up streams.
3 METHODOLOGY
UACANet combines a modified U-Net architecture, Parallel Axial Attention, and uncertainty-aware context modeling to refine polyp saliency predictions, especially around ambiguous regions.
- Overall Architecture: UACANet adds Parallel Axial Attention encoders and a decoder to a modified PraNet-style architecture for bottom-up and side-out feature processing.The PAA-e features support side-out fusion, PAA-d, and UACA while reducing input channels for these paths.
- Training Objective: Training uses binary cross-entropy and intersection-over-union losses across four predictions from PAA-d and UACA.The losses supervise predictions generated by the decoder and uncertainty-aware attention modules.
- Parallel Axial Attention: PAA computes horizontal- and vertical-axis non-local attention in parallel to capture global dependencies and local representations.The parallel arrangement gives both axes comparable contributions without trainable positional encoding.
- Uncertainty Augmented Context Attention: UACA derives foreground, background, and uncertain-area maps from a previous saliency prediction without requiring extra boundary guidance.The uncertain area is intended to capture ambiguity associated with boundary information.
- Uncertainty Augmented Context Attention: UACA aggregates input features with the three area maps into representative vectors, computes pixel-to-representation similarities, and forms a weighted context feature map.The context map is concatenated with the input feature map and passed through a point-wise convolution.
4 EXPERIMENTAL RESULTS
The experiments evaluate UACANet through implementation variants, ablations, dataset benchmarks, and qualitative comparisons. UACANet achieves strong results across five benchmarks, including its largest reported gain on challenging ETIS.
- Experimental Setup: Experiments cover implementation details, five benchmark datasets, ablations of PAA and uncertainty, and qualitative comparisons with prior methods.The evaluation uses mean Dice, mean intersection over union, and mean absolute error for ablation studies.
- Implementation Details: UACANet-S and UACANet-L use 32 and 256 channels outside the backbone, respectively, with Res2Net as the backbone.Images are resized to 352 × 352 for training and inference, with additional augmentation and 240 training epochs.
- Ablation Study: UACANet consistently outperforms CANet on mean Dice, mean intersection over union, and mean absolute error.The ablation compares models with and without the uncertainty map on CVC-ClinicDB and ETIS.
- Qualitative Analysis: UACANet’s uncertain area is closely related to polyp boundaries and produces more precise qualitative results than CANet.The qualitative comparison also reports detection of a polyp missed by the ground truth and detection of very small ETIS polyps.
- State-of-the-Art Comparison: 92.6% mean Dice on CVC-ClinicDB is a 2.7% improvement over PraNet, the latest state-of-the-art method.UACANet achieves the best performance among compared methods on CVC-ClinicDB and Kvasir.
- State-of-the-Art Comparison: 76.6% mean Dice on ETIS is a 13.8% improvement over PraNet on the most challenging benchmark.ETIS contains mostly small polyps that are hard to find, and UACANet detects small polyps that other methods miss qualitatively.
5 CONCLUSION
The paper concludes that UACANet augments context representations with uncertain areas for accurate polyp detection. It reports that uncertainty represents boundary information without expensive edge annotations and that the method outperforms prior state-of-the-art methods.
- Conclusion: UACANet augments uncertain areas in context representations to improve polyp detection.The network includes Parallel Axial Attention and Uncertainty Augmented Context Attention modules.
- Conclusion: Uncertain areas can represent complementary boundary information without requiring expensive edge annotations.The conclusion identifies uncertainty augmentation as complementary edge information in the context attention module.
- Conclusion: Quantitative and qualitative experiments show that UACANet outperforms previous state-of-the-art methods.