A Novel Computer-Aided Detection/Diagnosis System for Detection and Classification of Polyps in Colonoscopy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Colon polyp detection is crucial for colorectal cancer prevention.
• Current computer-aided detection systems face challenges with accuracy and miss rates.
• Deep learning offers potential for improving colon polyp detection and classification.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for colon polyp detection and classification.
• To enhance the performance of a You Only Look Once (YOLO) model using generative adversarial networks (GANs) and image deblurring techniques.
• To improve the accuracy and reduce the miss rate in colon polyp identification.

Main Methods:

• Utilized conditional GANs to generate synthetic polyp images, augmenting the training dataset for YOLO.
• Employed a data augmentation technique and a model trained with 300 GANs (GAN 300) for polyp detection.
• Applied Gaussian blurring to simulate colonoscopy image degradation and DeblurGAN-v2 for image deblurring prior to YOLO classification.

Main Results:

• The GAN 300 model achieved higher average precision (AP) compared to standard data augmentation for sessile serrated adenoma (SSA) and traditional adenoma (TA) detection.
• The GAN 300 model demonstrated significant improvements in AP, mean average precision (mAP), and intersection over union (IoU) for hyperplastic polyps (HP).
• DeblurGAN-v2 processing increased the mAP for polyp classification from 25.64% to 30.74%, indicating enhanced accuracy after deblurring.

Conclusions:

• Deep learning, particularly using GANs for data augmentation and deblurring techniques, effectively enhances colon polyp detection and classification accuracy.
• The developed system shows promise in reducing miss rates and improving the reliability of computer-aided diagnosis in colonoscopy.
• Further development of AI-powered tools can significantly aid clinicians in identifying and classifying colon polyps, potentially improving patient outcomes.