Abstract
Colorectal cancer is a significant global health threat, often originating from adenomatous polyps. While endoscopic examination is the primary screening method, high polyp miss rates remain a critical limitation. This study investigates the application of deep learning techniques to automate polyp localization and segmentation, aiming to improve detection accuracy and clinical efficiency. Using a combined dataset of 912 images (CVC-ColonDB and CVC-ClinicDB), we evaluated three convolutional neural network architectures: Fully Convolutional Network (FCN), SegNet, and a modified ResNet. Results indicate that deep learning approaches, particularly the modified ResNet, achieve high accuracy (>90%) and Intersection over Union (IoU) scores (>50%), demonstrating significant potential for real-time clinical assistance.
Introduction
Screening for colorectal cancer through colonoscopy is effective for both early detection and prevention. However, the procedure's efficacy is hampered by human factors, leading to a notable miss rate for polyps. Traditional computer-aided detection methods relied on hand-crafted features, which are often computationally expensive and struggle with the high morphological variability of polyps. This research explores deep learning as a more robust, time-saving alternative for automated segmentation.
Methods
The study framed polyp detection as a semantic segmentation task, classifying pixels as either "polyp" or "background." The researchers utilized a combined dataset of 912 images from the CVC-ColonDB and CVC-ClinicDB benchmarks.
Three distinct architectures were compared:
- FCN8: A standard for semantic segmentation that preserves spatial information by avoiding fully connected layers.
- SegNet: An encoder-decoder architecture optimized for pixel-wise efficiency.
- Modified ResNet: A Residual Network adapted with skip connections to facilitate deeper feature extraction and improve boundary definition.
Performance was measured using Accuracy and Intersection over Union (IoU), which quantifies the overlap between predicted and ground-truth masks.
Results
Quantitative analysis revealed that while all models achieved high overall accuracy, the modified ResNet was superior in precise segmentation. The results on the test set are summarized below:
| Architecture | IoU (Polyp) | IoU (Mean) | Accuracy (Mean) |
|---|---|---|---|
| FCN8 | 46.99% | 70.51% | 94.32% |
| SegNet | 53.80% | 74.35% | 95.18% |
| Modified ResNet | 64.03% | 80.03% | 96.29% |
Qualitative results (Figure 1) demonstrate that all three architectures successfully localized polyps. However, SegNet and the modified ResNet (rows b and c) produced significantly finer boundaries compared to FCN8.
Conclusion
Deep learning-based approaches offer a time-saving and effective solution for polyp segmentation in colonoscopy images compared to traditional hand-crafted methods. The evaluated architectures, particularly the modified ResNet, not only perform with high accuracy but also support near real-time processing. This suggests a strong potential for integrating these models into clinical workflows to assist endoscopists in reducing polyp miss rates.
Summary updated