Abstract
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. 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. Our model achieves state-of-the-art results across five public datasets, notably achieving a mean Dice Similarity Coefficient (DSC) of 0.878 on a combined dataset—a 14% improvement over previous benchmarks. This study demonstrates the potential for deep learning to provide fast and accurate polyp segmentation for clinical use.
Introduction
Colorectal cancer (CRC) is a leading cause of cancer-related mortality globally, yet it is highly preventable and treatable if detected early. While colonoscopy is the gold-standard screening tool, its effectiveness is limited by human factors; clinicians miss between 2% and 26% of polyps during procedures. These missed lesions are often small, flat, or obscured by mucosal folds.
Computer-aided detection (CADe) and diagnosis (CADx) systems, specifically those utilizing semantic segmentation, offer a solution by highlighting polyps in real-time. However, developing these systems is challenging due to the high variability in polyp shape, color, and texture, as well as the "class imbalance" problem where the polyp occupies only a small fraction of the endoscopic image. This paper introduces the Focus U-Net to address these challenges through architectural innovations in attention and loss functions.
Methodology
The Focus U-Net Architecture
The Focus U-Net builds upon the standard U-Net architecture, a staple in medical image segmentation, by introducing several key enhancements designed for high-precision polyp detection:
- The Focus Gate (FG): Standard attention mechanisms in medical imaging typically focus on spatial attention (the "where"). The Focus Gate introduces a dual-pathway approach, combining spatial and channel-based attention (the "what"). This allows the model to simultaneously localize a polyp and identify its specific features, such as unique textures or colors that distinguish it from healthy tissue.
- The Focal Parameter ($\gamma$): The authors integrated a tunable focal parameter directly into the attention mechanism. Acting much like a contrast control, this parameter allows the network to aggressively suppress background noise and sharpen the focus on the polyp's boundaries.
- Deep Supervision and Short-Range Skips: To ensure robust feature learning, the model utilizes deep supervision, generating predictions at multiple layers. This prevents the "vanishing gradient" problem and ensures that features at different scales are preserved throughout the network.
Hybrid Focal Loss (HFL)
To combat the class imbalance inherent in endoscopic images, the researchers developed the Hybrid Focal Loss. By combining Focal Loss (which focuses on hard-to-classify pixels) and Focal Tversky Loss (which balances precision and recall), the model is forced to prioritize the detection of small or difficult polyps that might otherwise be ignored by standard loss functions.
Experimental Setup
The model was evaluated using five diverse public datasets to ensure robustness across different endoscopic equipment and patient populations:
- CVC-ClinicDB and Kvasir-SEG: The primary benchmarks for polyp segmentation.
- CVC-ColonDB, ETIS-Larib PolypDB, and EndoScene: Challenging datasets characterized by varied lighting and smaller, more difficult-to-see polyps.
- Combined Dataset: A rigorous evaluation using all five datasets to test the model's ability to generalize to unseen data.
Performance was measured using the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU).
Results and Discussion
The Focus U-Net established new state-of-the-art (SOTA) benchmarks across all tested datasets. On the standard CVC-ClinicDB and Kvasir-SEG datasets, the model achieved mean DSC scores of 0.941 and 0.910, respectively.
The most significant advancement was observed in the combined five-dataset evaluation. The Focus U-Net achieved a mean DSC of 0.878 and a mean IoU of 0.809. This represents a 14% and 15% improvement, respectively, over the previous SOTA (HarDNet-MSEG). These results suggest that the dual-attention mechanism and the focal parameter are highly effective at capturing the nuances of polyp morphology across varied clinical environments.
Clinical Implications and Limitations
Despite its high accuracy and efficiency—making it suitable for potential real-time application—the model faces one primary hurdle: the "all-polyp" training bias. Because the training datasets consisted exclusively of images containing polyps, the model may produce false positives when encountering the "normal" (polyp-free) frames that constitute the majority of a real-world colonoscopy video.
Conclusion
The Focus U-Net represents a significant step forward in automated polyp segmentation. By combining spatial and channel attention with a novel hybrid loss function, the model provides a robust framework for reducing polyp miss rates. Future research will focus on incorporating "negative" (healthy) samples into training to improve specificity and testing the architecture in live clinical pipelines.
Summary updated