New polyp image classification technique using transfer learning of network-in-network structure in endoscopic images

Area of Science:

• Medical Imaging and Artificial Intelligence
• Gastroenterology and Oncology

Background:

• Colorectal cancer is a leading global cancer, with polyps increasing risk.
• Early detection and resection of colorectal polyps are crucial for cancer prevention.
• Accurate polyp classification is essential for effective endoscopic procedures.

Purpose of the Study:

• To improve the performance of colorectal polyp classification using deep learning.
• To develop an automated algorithm assisting endoscopists in identifying adenomatous polyps.

Main Methods:

• Fine-tuning a pre-trained Network-in-Network (NIN) model on 1000 colonoscopy images.
• Utilizing transfer learning from ImageNet and comparing with AlexNet-based methods.
• Evaluating performance through 20 experiments with 800 training and 200 testing images per set.

Main Results:

• The proposed NIN method demonstrated statistically significant higher accuracy than four state-of-the-art methods.
• An 18.9% improvement in accuracy was observed compared to a standard AlexNet approach.
• High performance metrics achieved: Area Under the Curve (AUC) of ~0.930 ± 0.020 and recall rate of ~0.929 ± 0.029.

Conclusions:

• The enhanced AI algorithm significantly improves polyp classification accuracy and recall.
• This system can assist endoscopists in timely identification and resection of precancerous polyps.
• Automated polyp classification facilitates early-stage cancer prevention and improved patient outcomes.