Abstract
Colorectal cancer is a major global health concern, with the majority of cases originating from intestinal polyps. Early screening via Wireless Capsule Endoscopy (WCE) is vital but labor-intensive. This study proposes a Deep Convolutional Neural Network (DCNN) for automated polyp segmentation. Using the CVC-ColonDB and CVC-ClinicDB datasets, the model achieved an Intersection over Union (IoU) of over 0.5 on test data. The results demonstrate the potential of DCNNs to assist clinicians by providing automated visualization of polyps without manual feature extraction.
Introduction
Colorectal cancer ranks as the third most common cancer in men and the second in women. Since 80%–95% of these cancers evolve from intestinal polyps, early detection is the most effective prevention strategy. Wireless Capsule Endoscopy (WCE) provides a non-invasive means to examine the digestive tract; however, the sheer volume of images generated per patient creates a significant diagnostic burden. Computer-assisted segmentation offers a solution to help doctors visualize and identify polyps more efficiently.
Methods
The researchers implemented a Deep Convolutional Neural Network (DCNN) for the segmentation task. DCNNs are particularly effective for image segmentation as they learn hierarchical features directly from raw data, eliminating the need for manual preprocessing or "hand-crafted" feature extraction.
Data and Training
The model was trained and evaluated using two open-access datasets: CVC-ColonDB and CVC-ClinicDB. The data was split into training, validation, and test sets.
- Loss Function: Cross-entropy loss was used for optimization.
- Evaluation Metric: Performance was measured using the Intersection over Union (IoU) value, which quantifies the overlap between predicted masks and ground truth annotations.
Results
The quantitative evaluation shows that the model performs well across all subsets, though a performance gap exists between training and test data, likely due to limited data and high inter-patient variance.
| Dataset | Background IoU | Polyp IoU | Accuracy |
|---|---|---|---|
| Training | 0.976 | 0.800 | 0.987 |
| Validation | 0.969 | 0.604 | 0.970 |
| Test | 0.947 | 0.523 | 0.725 |
As shown in the visual results in Figure 2, the model successfully identifies the polyp regions across various frames, producing segmentation masks that align closely with the ground truth.
Conclusion
The experimental results confirm that the DCNN model is capable of effective polyp segmentation, achieving an IoU greater than 0.5 on unseen test data. A key advantage of this approach is its adaptability to polyps of different sizes and morphologies without requiring manual feature engineering.
Future work will focus on expanding the training datasets to improve generalization and exploring more advanced architectures to refine segmentation boundaries and improve performance on diverse clinical data.
Summary updated