MKDCNet

MKDCNet segments polyps in colonoscopy images to improve polyp delineation and support early detection and prevention of colorectal cancer.

Key Features:

• Architecture: An encoder-decoder neural network that leverages a pre-trained ResNet50 as its encoder for feature extraction from colonoscopy images.
• Multiple Kernel Dilated Convolution (MKDC) Block: Expands the receptive field using multiple kernel dilated convolutions to capture heterogeneous and robust representations of polyp features.
• Robustness Across Datasets: Evaluated on four publicly available polyp datasets and a cell nuclei dataset, showing performance comparable to or better than existing methods on same-distribution and unseen datasets.
• Efficiency: Processes images at approximately 45 frames per second on an RTX 3090 GPU, supporting real-time inference.

Scientific Applications:

• Real-time clinical colonoscopy systems: Enables live polyp segmentation during colonoscopy to assist detection workflows.
• Computer-aided diagnosis (CAD): Provides consistent polyp delineation to reduce variability among endoscopists and lower miss rates of colorectal polyps.
• Cross-dataset evaluation: Applicable for benchmarking segmentation performance on diverse polyp and cell nuclei imaging datasets.

Methodology:

MKDCNet uses a pre-trained ResNet50 as the encoder within an encoder-decoder architecture; extracted features are processed by the MKDC block, which expands the receptive field via multiple kernel dilated convolutions to learn diverse representations of polyps.