Dual-Branch Multi-Task Regressor and Transformer Model for Endoscopic Image Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Endoscopic diagnosis of colon cancer is critical for early detection.
• Manual image analysis is time-consuming and requires expert interpretation.
• Automated image classification offers a valuable solution to improve efficiency.

Purpose of the Study:

• To develop a novel multi-label classification method for endoscopic images.
• To integrate local and global feature learning for enhanced classification accuracy.
• To improve the early diagnosis of colon cancer through automated image analysis.

Main Methods:

• A hybrid model combining a Swin Transformer (global features) and a modified VGG16-CNN (local features).
• Incorporation of saliency maps and texture features within a multi-task learning framework to boost CNN performance.
• Utilized the Kvasir-v2 dataset for training and evaluation.

Main Results:

• The proposed model achieved a 96.08% F1-score and 96.06% accuracy.
• Outperformed existing state-of-the-art methods in endoscopic image classification.
• Demonstrated superior performance in identifying colon cancer from endoscopic visuals.

Conclusions:

• The developed AI model shows significant potential for clinical application in colon cancer diagnosis.
• The integration of local and global features enhances classification accuracy.
• This approach can lead to more efficient and accurate early detection of colon cancer.