LPSNet: A Novel Lightweight Polyp Segmentation Convolutional Neural Network

Colorectal cancer is a leading cause of cancer-related deaths worldwide, making early detection essential for improving patient outcomes. Polyp segmentation in colonoscopy images is vital in identifying potentially malignant lesions. This paper presents LPSNet, a lightweight convolutional neural network designed for efficient polyp segmentation. It employs a sequential architecture with skip connections, ensuring effective feature extraction and knowledge transfer. The network comprises two convolution-convolution-pooling blocks for encoding, two convolution-convolution-transposed convolution blocks for feature extraction, and a convolution-convolution-convolution block for decoding the output mask. LPSNet effectively retains fine-grained features while preserving high-level contextual information by integrating short and long skip connections. It is evaluated using the publicly available CVC-ClinicDB and Kvasir datasets. Extensive comparisons with existing models demonstrated that LPSNet outperformed previous approaches in segmentation accuracy. On the Kvasir dataset, LPSNet achieved a mean absolute error of 0.0234, a mean intersection over union of 0.8693, and a mean Dice score of 0.9207. A similar evaluation of the CVC-ClinicDB dataset achieved 0.0081, 0.8714, and 0.9241. The computational complexity of the proposed LPSNet is the lowest among the compared models, indicating its potential as a reliable and efficient solution for polyp segmentation and making it particularly well-suited for real-time medical applications in the future.