Source-linked AI summary
Focus U-Net: A novel dual attention-gated CNN for polyp segmentation during colonoscopy
Michael Yeung, Evis Sala, Carola-Bibiane Schönlieb, Leonardo Rundo
TL;DR
Colonoscopy is effective but can miss polyps, particularly when multiple or small adenomas are present. This paper introduces Focus U-Net, which combines dual attention gating, skip-connection design, deep supervision, and Hybrid Focal loss for polyp segmentation. Across five public datasets, it reports state-of-the-art performance, including mean DSC 0.878 and mean IoU 0.809.
Problem
Colonoscopy has significant polyp miss rates, creating an opportunity for computer-aided systems to support clinicians and reduce missed polyps.
Method
Focus U-Net combines spatial and channel attention in a Focus Gate with short-range skip connections, deep supervision, and Hybrid Focal loss.
Results
0.878 mean DSC and 0.809 mean IoU were achieved on a combination of five public polyp datasets, improving over previous state-of-the-art results by 14% and 15%, respectively.
Takeaways & Limitations
The study shows the potential of deep learning to provide fast and accurate polyp segmentation during colonoscopy.
Takeaways & Limitations
The study reports limitations associated with its current design, including an author-identified limitation in the experimental results and discussion.
Abstract
from arXiv · showhide
Background: Colonoscopy remains the gold-standard screening for colorectal cancer. However, significant miss rates for polyps have been reported, particularly when there are multiple small adenomas. This presents an opportunity to leverage computer-aided systems to support clinicians and reduce the number of polyps missed. Method: In this work we introduce the Focus U-Net, a novel dual attention-gated deep neural network, which combines efficient spatial and channel-based attention into a single Focus Gate module to encourage selective learning of polyp features. The Focus U-Net further incorporates short-range skip connections and deep supervision. Furthermore, we introduce the Hybrid Focal loss, a new compound loss function based on the Focal loss and Focal Tversky loss, to handle class-imbalanced image segmentation. For our experiments, we selected five public datasets containing images of polyps obtained during optical colonoscopy: CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB, ETIS-Larib PolypDB and EndoScene test set. To evaluate model performance, we use the Dice similarity coefficient (DSC) and Intersection over Union (IoU) metrics. Results: Our model achieves state-of-the-art results for both CVC-ClinicDB and Kvasir-SEG, with a mean DSC of 0.941 and 0.910, respectively. When evaluated on a combination of five public polyp datasets, our model similarly achieves state-of-the-art results with a mean DSC of 0.878 and mean IoU of 0.809, a 14% and 15% improvement over the previous state-of-the-art results of 0.768 and 0.702, respectively. Conclusions: This study shows the potential for deep learning to provide fast and accurate polyp segmentation results for use during colonoscopy. The Focus U-Net may be adapted for future use in newer non-invasive screening and more broadly to other biomedical image segmentation tasks involving class imbalance and requiring efficiency.
1. Introduction
Colonoscopy is the gold-standard CRC screening tool but has substantial polyp miss rates, especially for patients with multiple or small adenomas. These detection challenges motivate computer-aided systems that can segment polyp boundaries accurately.
- Colonoscopy remains the gold-standard CRC screening tool, offering high sensitivity and specificity.
- Polyp miss rates increase with multiple polyps and vary from 2% for adenomas ≥10mm to 26% for adenomas < 5mm.
- Computer-aided systems could reduce polyp miss rates during colonoscopy.
- Polyp segmentation aims to delineate polyp borders from surrounding mucosa when polyps enter the operator’s field of view but remain undetected.
- Hand-crafted features produced low segmentation accuracy because they could not capture the scale of polyp heterogeneity.
2. Related work
Prior polyp-segmentation models achieved strong results but often imposed memory and inference-time costs that limit clinical use. The paper therefore introduces Focus U-Net, combining attention mechanisms with an efficiency-oriented design for accurate segmentation.
- U-Net-based models dominate polyp segmentation, using encoder context and decoder localization to produce segmentation maps.
- Attention-based models incorporate spatial and channel attention, while other methods use transfer learning, progressive boundary refinement, or low-memory backbones.
- Large memory requirements and long inference times limit some high-performing models in clinical practice, where real-time segmentation is required.
- Focus U-Net introduces a Focus Gate combining spatial and channel-based attention with a focal parameter controlling background suppression.
- The model achieves state-of-the-art results across five public polyp-segmentation datasets.
3. The proposed Focus U-Net architecture
Focus U-Net uses attention-gated skip connections to selectively emphasize polyp-related regions and features during decoding. Its architecture also adds short-range skip connections, deep supervision, and a compound Hybrid Focal loss.
- The encoder captures polyp-relevant features, while the decoder progressively upsamples them to produce a precise polyp segmentation map.
- The deepest encoder layer supplies a gating signal that refines features from long-range skip connections before decoder fusion.
- Short-range skip connections and deep supervision create additional information pathways and shortcuts for backward loss propagation.
- Attention Gates use contextual gating information to prune skip connections, highlight regions of interest, and reduce false-positive predictions.
- The Focus Gate combines spatial and channel attention, using pooling and adaptive convolution to generate attention coefficients for feature weighting.
- Hybrid Focal loss is defined as the sum of Focal Tversky loss and Focal loss.
4. Materials and evaluation methods
The study evaluates Focus U-Net for polyp segmentation using five public colonoscopy datasets, standardized preprocessing, ablation protocols, and DSC and IoU metrics.
- Datasets: Five public datasets provide colonoscopy polyp images for evaluating Focus U-Net segmentation.The datasets are CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB, ETIS-Larib PolypDB, and the EndoScene test set.
- Datasets: The datasets vary in size, resolution, provenance, and patient or video-sequence composition.Examples include 612 CVC-ClinicDB frames from 23 sequences and 1,000 Kvasir-SEG images with resolutions ranging from 332×487 to 1920×1072 pixels.
- Evaluation protocol: Ablations use five-fold cross-validation on CVC-ClinicDB to add the Focus Gate, Hybrid Focal loss, short-range skips, and deep supervision sequentially.Baseline comparisons include U-Net and Attention U-Net, with models trained for 100 epochs using stochastic gradient descent with Nesterov momentum.
- Metrics: Segmentation accuracy is assessed primarily with DSC and IoU, which provide complementary overlap-based perspectives.IoU penalizes isolated poor pixel classifications more heavily than DSC, while recall and precision add information about false negatives and false positives.
5. Experimental results
Focus U-Net improves segmentation performance across individual and combined polyp datasets, with gains over established baselines and consistent qualitative results across five datasets.
- Ablation study: 0.875 ± 0.016 DSC is achieved by Focus U-Net in ablation, versus 0.828 ± 0.021 for U-Net and 0.801 ± 0.019 for Attention U-Net.The ablation progressively adds the model components and reports performance gains with successive additions.
- CVC-ClinicDB: 0.941 mDSC and 0.893 mIoU are achieved on CVC-ClinicDB, establishing state-of-the-art results.Focus U-Net also provides the best recall-precision balance in this comparison.
- Kvasir-SEG: 0.910 mDSC and 0.845 mIoU are achieved on Kvasir-SEG, with the highest mIoU reported in that comparison.The results are presented as state-of-the-art performance for the dataset.
- Qualitative results: Focus U-Net produces consistently accurate segmentations across five public datasets, while poorest cases involve difficult polyps or poor-quality images.The intermediate predictions also indicate that deeper layers can localize polyps effectively.
6. Discussion and conclusion
The Focus U-Net is a lightweight attention-gated architecture designed for efficient polyp segmentation, achieving strong and consistent performance across five public datasets. Its broader applicability is suggested for efficient, class-imbalanced biomedical segmentation, but clinical deployment remains to be established.
- Architecture and optimisation: The Focus U-Net combines a Focus Gate with short-range skip connections, deep supervision, and Hybrid Focal loss.The Focus Gate combines efficient channel and spatial attention, while Hybrid Focal loss addresses class imbalance.
- Performance: The model achieves state-of-the-art performance across five public polyp datasets, with consistent generalisation to unseen data.The authors report that the architecture performs consistently well across datasets and generalises to data from different datasets.
- Clinical-use considerations: The architecture is lightweight and designed for fast inference during live colonoscopy.Efficient channel and spatial attention mechanisms and a lightweight U-Net backbone support the real-time setting.
- Performance: 0.878 mean DSC and 0.809 mean IoU were achieved on the combined five-dataset evaluation.These results represent 14% and 15% improvements over previous state-of-the-art values of 0.768 and 0.702.
- Clinical-use considerations: Segmentation-based detection can provide operator-independent polyp-size estimates that may guide biopsy decisions.The authors identify this as a secondary advantage of segmentation-based computer-aided detection.
- Limitations: The study is limited because training datasets contain polyps in every image, unlike most live video data, potentially yielding a higher false-positive rate.The authors state that training with images without polyps can produce poorer generalisation and that false positives are less undesirable than false negatives for this system's purpose.