Localisation of Colorectal Polyps by Convolutional Neural Network Features Learnt from White Light and Narrow Band...

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Accurate localization of colorectal polyps is crucial for effective diagnosis and treatment.
• Existing polyp localization algorithms often suffer from limited data diversity and potential overfitting due to training and testing on the same datasets.

Purpose of the Study:

• To develop and evaluate a unified, real-time object detection system using a You-Only-Look-Once (YOLO) convolutional neural network (CNN) for localizing colorectal polyps in endoscopic images.
• To assess the performance of the YOLO model trained on diverse datasets, including novel narrow-band (NB) imaging data.

Main Methods:

• A You-Only-Look-Once (YOLO) convolutional neural network (CNN) was employed for polyp localization.
• The YOLO model underwent pre-training on non-medical images and subsequent fine-tuning using colonoscopic images from three distinct databases, including a newly collected set from 106 patients using narrow-band (NB) imaging.
• Performance was evaluated on an independent public database of 196 white light (WL) images.

Main Results:

• The YOLO model achieved a precision of 79.3% and a sensitivity of 68.3% for polyp localization when trained on augmented images from multiple WL databases.
• The system demonstrated high time efficiency, processing at 0.06 seconds per frame.
• Distinct differences in CNN features between white light (WL) and narrow-band (NB) endoscopic images were observed, suggesting the need for modality-specific considerations.

Conclusions:

• The YOLO algorithm demonstrates significant potential as an assistive tool for endoscopists in localizing colorectal polyps during endoscopic procedures.
• Future research should focus on creating large-scale, diverse databases encompassing various imaging modalities and scenarios to enhance model generalizability and clinical applicability.