IDDF2018-ABS-0257 Detecting and segmenting polyps using a deep learning-based model

Liansheng Wang, Shuxin Wang, Yanxing Hu, Shaohui HuangClinical 2018 · DOI

Introduction

Colorectal cancer is the third most common cancer globally, typically developing from untreated polyps. Early detection and removal of these polyps are critical for prevention. Wireless Capsule Endoscopy (WCE) provides a non-invasive method for visualizing the intestinal tract without surgery. However, a single WCE procedure generates approximately 25,000 images per patient. Manually reviewing this volume of data is extremely time-consuming and exhausting for clinicians, increasing the risk of missing abnormalities. This study presents a computer-aided diagnosis (CAD) system designed to automate the detection and segmentation of polyps in WCE images.

Methods

The proposed system utilizes a deep learning-based model, specifically the U-Net architecture. U-Net is an encoder-decoder model: the encoder captures high-level context and features, while the decoder enables precise localization by projecting these features back to the original image resolution.

To optimize the model for medical image segmentation, the researchers employed a hybrid loss function combining Dice loss and cross-entropy loss. This approach balances pixel-wise classification accuracy with the need to maximize the overlap between predicted masks and ground truth annotations, which is particularly important given the small size of polyps relative to the background.

Datasets

The model was trained and evaluated using images from the CVC-Colon and CVC-Clinic datasets. A total of 992 images containing polyps were used. The data was partitioned into training, validation, and test sets to ensure robust evaluation.

Table 1: Distribution of training, validation, and test data across the CVC-Colon and CVC-Clinic datasets.

Results

The performance of the polyp segmentation was primarily evaluated using the Dice coefficient, which measures the similarity between the model's prediction and the expert-defined ground truth. The system achieved a Dice coefficient of 84.15%.

Example of a model prediction mask identifying polyp regions within a WCE frame.

The results indicate that the U-Net model can automatically learn relevant features and efficiently segment polyps of various sizes. This automation significantly reduces the manual workload for physicians, allowing them to focus on confirmed abnormalities rather than scanning thousands of healthy frames.

Conclusion

WCE is a vital tool for colorectal cancer screening, but the data volume presents a significant clinical bottleneck. This study demonstrates that a deep learning-based CAD system using U-Net can effectively segment polyps with high accuracy. Future work will focus on clinical trials to further validate the system's impact on reducing physician fatigue and improving diagnostic efficiency in real-world settings.

Summary updated

Abstract

<h3>Background</h3> Colorectal cancer is the third most common cancer in the world, which developed from untreated polyps. Detecting polyps in their early stage have become a serious medical issue. Wireless capsule endoscopy (WCE) was designed to examine the intestinal diseases without surgery and can give a direct visualisation of intestines, in which the time cost is to analyse lots of images to capture the abnormal parts. The doctors will be wearied with about 25 000 images per person produced by WCE. This study aims to build a computer-aided diagnosis system to help doctors analyse WCE images and detect polyps. <h3>Methods</h3> We use a deep learning-based model to build a computer-aided diagnosis system of WCE images for the assignment of detecting and segmenting polyps. The U-Net network with both dices loss and cross entropy loss is employed in this work, which is considered as an encoding and decoding model. The data used were available from CVC-Colon and CVC-Clinic, it has 992 images with polyps totally. The data is shown in table 1. Comparison to other algorithms, the U-Net model can learn more features automatically and is a high-efficient method for our assignment. <h3>Results</h3> The performance of polyp segmentation was evaluated by Dice’s coefficient (also known as the Dice coefficient), which presents the degree of similarity between prediction and ground truth. The Dice’s coefficient in our study is 84.15%. The segmentation results are shown in IDDF2018-ABS-0257 figure 1. <h3>Conclusions</h3> WCE plays an essential role in diagnosis and prevention of colorectal cancer. However, distinguishing abnormal WCE images with polyps is a significant challenge because it needs high time cost. The results show that our method can help physicians analyse the WCE images and reduce their pressure. This method could be further utilised in the clinical trials to help physicians from the tedious image analysing work.
BibTeX
@inproceedings{wang2018iddf,
  author = {Liansheng Wang and Shuxin Wang and Yanxing Hu and Shaohui Huang},
  title = {{IDDF2018-ABS-0257 Detecting and segmenting polyps using a deep learning-based model}},
  booktitle = {Clinical},
  year = {2018},
  doi = {10.1136/gutjnl-2018-iddfabstracts.178},
}
Figure