Toward real-time polyp detection using fully CNNs for 2D Gaussian shapes prediction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonoscopy is crucial for detecting colon polyps, but miss-rates remain a challenge.
• Accurate real-time polyp detection systems are needed to improve colonoscopy efficacy.
• Current fully convolutional neural networks (F-CNNs) for polyp detection often struggle with accuracy and false positives.

Purpose of the Study:

• To develop an accurate and reliable real-time colon polyp detection system.
• To enhance the performance of F-CNNs by proposing a novel mask training approach.
• To reduce the miss-rate of colon polyps during colonoscopy procedures.

Main Methods:

• Utilized single-shot feed-forward fully convolutional neural networks (F-CNNs).
• Proposed the use of 2D Gaussian masks for training object detection models, instead of traditional binary masks.
• Evaluated the system on two benchmark colon polyp datasets: ETIS-LARIB and CVC-ColonDB.

Main Results:

• The proposed 2D Gaussian masks demonstrated efficiency in detecting flat and small polyps with unclear boundaries.
• The novel mask approach improved the discrimination of polyps from false positives, enhancing training effectiveness.
• Achieved state-of-the-art results, including 86.54% recall and 86.12% precision on the ETIS-LARIB dataset, and 91% recall and 88.35% precision on the CVC-ColonDB dataset.

Conclusions:

• 2D Gaussian masks offer a significant improvement for real-time polyp detection using F-CNNs.
• The proposed method effectively reduces false positives and enhances the detection of challenging polyp types.
• This approach holds promise for improving the accuracy and reliability of colonoscopy by reducing polyp miss-rates.