Multiclassification of Endoscopic Colonoscopy Images Based on Deep Transfer Learning

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Rising incidence of colorectal cancer necessitates improved diagnostic tools.
• Current deep learning applications in colonoscopy assist polyp detection but struggle with accurate lesion classification.
• Distinguishing between polyps and other intestinal conditions remains a challenge.

Purpose of the Study:

• To develop a deep learning-based multiclassification method for medical colonoscopy images.
• To improve the accuracy of classifying intestinal conditions, specifically differentiating polyps from inflammation, tumors, and normal tissue.
• To enhance the recognition rate of polyps in colonoscopy images.

Main Methods:

• Utilized transfer learning with a pre-trained network on ImageNet as a base model.
• Fine-tuned the pre-trained model using a custom dataset of medical colonoscopy images.
• Applied the fine-tuned model for multiclassification of polyps, inflammation, tumors, and normal intestinal conditions.

Main Results:

• The proposed method significantly improved the recognition rate of polyps.
• Classification accuracy for other categories (inflammation, tumor, normal) was maintained.
• Demonstrated the effectiveness of transfer learning and fine-tuning for medical image classification with limited data.

Conclusions:

• The deep learning multiclassification method effectively aids in diagnosing various intestinal diseases from colonoscopy images.
• The approach assists physicians by improving polyp recognition and overall classification accuracy.
• This technology holds potential for earlier and more accurate detection of colorectal conditions.