Automated polyp segmentation based on a multi-distance feature dissimilarity-guided fully convolutional network.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Gastroenterology and Oncology

Background:

• Colorectal polyps, precursors to malignancy, are often detected via colonoscopy.
• Accurate polyp segmentation is challenging due to their variable characteristics and surrounding structures.
• Existing convolutional neural network (CNN) models struggle with small polyps and contextual similarities, leading to missed or false detections.

Purpose of the Study:

• To develop a novel, accurate automatic polyp segmentation approach.
• To address limitations of current CNN models in capturing polyp features and details.

Main Methods:

• Introduction of a multi-distance feature dissimilarity-guided fully convolutional network.
• Incorporation of a multi-distance difference (MDD) module using multi-layer feature subtraction (MLFS) for enhanced feature extraction.
• Utilization of a hybrid loss (HL) module to supervise feature maps and improve prediction accuracy.

Main Results:

• The proposed approach demonstrated superior performance in automatic polyp segmentation.
• Outperformed five state-of-the-art methods across six evaluation criteria on four datasets.
• Effectively captured discriminative features and complementary information across different network layers.

Conclusions:

• The novel network effectively segments colorectal polyps, overcoming challenges posed by polyp variability and contextual structures.
• The MDD and HL modules significantly contribute to improved feature representation and prediction accuracy.
• This approach holds promise for enhancing automated polyp detection in colonoscopy, aiding early cancer diagnosis.