A novel Joint-Net model for recognizing small-bowel polyp images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate polyp detection in enteroscopy images is crucial for preventing pathological changes.
• Automated recognition systems can improve diagnostic efficiency and accuracy.

Purpose of the Study:

• To propose a novel deep learning model, Joint-Net, for automatic polyp recognition in enteroscopy images.
• To evaluate the performance of Joint-Net in segmenting and classifying polyps.

Main Methods:

• The Joint-Net architecture combines transfer learning with VGG16 and a U-Net-based structure.
• Modifications include added convolution layers, identity blocks in skip connections, and an asymmetric convolution layer for connecting network halves.
• A loophole-like structure is employed in the output layer.

Main Results:

• The model achieved a mean Dice coefficient of 90.05% and an intersection over union (IoU) of 82.71% for polyp segmentation.
• Classification accuracy for distinguishing normal and polyp images reached 93.50%.

Conclusions:

• The developed Joint-Net model demonstrates successful segmentation and recognition of polyps in enteroscopy images.
• This approach shows promise for automated polyp detection in clinical settings.