Application of Convolutional Neural Networks for Automated Ulcer Detection in Wireless Capsule Endoscopy Images

Haya Alaskar, Abir Hussain, Nourah Alaseem, Panos Liatsis, Dhiya Al-Jumeily

TL;DR

This paper introduces an automated system using Convolutional Neural Networks (CNNs), specifically GoogLeNet and AlexNet, for detecting ulcers in Wireless Capsule Endoscopy (WCE) images. By fine-tuning pre-trained CNNs, the research achieves high accuracy in identifying ulcers, offering a promising automated diagnostic tool to assist clinicians in analyzing WCE data.

What Problem Does This Paper Address?

Wireless Capsule Endoscopy (WCE) generates a massive number of images (over 55,000 per patient) that physicians must meticulously review to detect abnormalities like ulcers. This manual process is time-consuming and prone to errors due to challenges inherent in WCE images, such as low contrast, complex backgrounds, and variations in lesion appearance. Existing traditional machine learning methods often require extensive manual feature engineering, which can be suboptimal. This paper aims to address this by developing an automated system that can accurately and efficiently detect ulcers in WCE images, thereby reducing the burden on clinicians and improving diagnostic speed and reliability.

Methods

The core of this research lies in the application and fine-tuning of two pre-trained Convolutional Neural Network (CNN) architectures: GoogLeNet and AlexNet. These networks were chosen for their proven success in image recognition tasks.

  • Pre-trained CNNs: Instead of training CNNs from scratch, which requires vast amounts of data and computational resources, the researchers utilized pre-trained models. These models were initially trained on a large dataset (ImageNet) and then "fine-tuned" for the specific task of ulcer detection.
  • Fine-tuning: This involved freezing the weights of the initial layers of the CNNs (which learn general features like edges and textures) and retraining the later, fully connected layers. These later layers are responsible for mapping the extracted features to the specific classes (ulcer or normal).
  • Architecture Modifications: Minor adjustments were made to the architectures of GoogLeNet and AlexNet to suit the ulcer detection task, such as adding dropout layers for regularization and modifying the final classification layers to output two classes (normal and abnormal/ulcer).
  • Evaluation Metrics: The performance of the CNNs was assessed using standard metrics: accuracy, sensitivity, specificity, loss, and Area Under the Curve (AUC).

Datasets

The study utilized a dataset of 1875 WCE images, comprising 1525 ulcer instances and 250 normal class instances. These images were captured from both esophageal and gastric regions of the digestive tract. The dataset was split into training (80%) and testing (20%) sets. The original image resolution was 256x256x3 pixels. For GoogLeNet, images were resized to 224x224x3, and for AlexNet, to 227x227x3 pixels.

Key Insights and Ideas

  • Pre-trained CNNs are powerful for limited datasets: The research demonstrates that fine-tuning pre-trained CNNs is a highly effective strategy for medical image analysis tasks where large, labeled datasets are scarce. This approach leverages the general feature extraction capabilities learned from massive datasets.
  • Automated feature extraction is key: CNNs excel at automatically extracting relevant features from images, eliminating the need for laborious manual feature engineering that is characteristic of traditional machine learning methods. This is crucial for complex medical images with subtle abnormalities.
  • Learning rate significantly impacts performance: The study highlights that the choice of learning rate is critical for achieving optimal performance. Very high learning rates led to unsatisfactory results, while lower rates, particularly 0.0001 and 0.001, resulted in high accuracy and stability.
  • CNNs offer a significant leap over traditional methods: The results show that CNNs achieve superior performance compared to traditional machine learning techniques, often by large margins, underscoring their potential as advanced diagnostic tools.

Key Results

  • High Accuracy: Both GoogLeNet and AlexNet achieved excellent performance. With optimal parameter settings (specifically, a learning rate of 0.0001), both models demonstrated 100% accuracy in detecting ulcers. AlexNet also achieved 100% accuracy with a learning rate of 0.001.
  • Sensitivity and Specificity: The models achieved 100% sensitivity and specificity, indicating they could correctly identify both abnormal (ulcer) and normal instances with high confidence.
  • Runtime Efficiency: AlexNet was significantly faster than GoogLeNet in terms of training time, requiring approximately half the time (18:37 min vs. 37:41 min) for training on the same dataset. Both CNNs, however, were considerably faster than traditional methods that involve extensive manual feature extraction and selection.
  • ROC Curves: Receiver Operating Characteristic (ROC) curves illustrated a balanced trade-off between sensitivity and specificity for both CNNs, further validating their classification performance.

Limitations and Future Work

  • Dataset Size: While the study achieved impressive results, the authors acknowledge that the dataset, though substantial for WCE research, is still relatively limited.
  • Generalizability: The performance of the models might be influenced by the specific characteristics of the dataset. Further validation on larger and more diverse datasets from different sources would be beneficial.
  • Other Abnormalities: The current study focuses solely on ulcer detection. Future work could extend this approach to detect other types of abnormalities present in WCE images, such as bleeding, polyps, or inflammatory lesions.
  • Real-time Application: While the CNNs are efficient, further optimization might be needed for true real-time analysis during a live endoscopy procedure.
  • Interpretability: Like many deep learning models, understanding why the CNN makes a particular prediction (explainability) remains an area for future research.

Summary updated

BibTeX
@inproceedings{alaskar2019application,
  author = {Haya Alaskar and Abir Hussain and Nourah Alaseem and Panos Liatsis and Dhiya Al-Jumeily},
  title = {{Application of Convolutional Neural Networks for Automated Ulcer Detection in Wireless Capsule Endoscopy Images}},
  booktitle = {Sensors},
  year = {2019},
  url = {https://openreview.net/forum?id=FJBhXEvboA},
}
Figure