Li-SegPNet: Encoder-Decoder Mode Lightweight Segmentation Network for Colorectal Polyps Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Colorectal cancer diagnosis relies on accurate segmentation of gastrointestinal lesions, particularly polyps.
• Automated polyp segmentation is a challenging but critical task for early cancer detection.

Purpose of the Study:

• To present a novel, lightweight encoder-decoder architecture, Li-SegPNet, with an attention mechanism for polyp segmentation.
• To improve the accuracy and efficiency of automated polyp detection in colonoscopy images.

Main Methods:

• Developed Li-SegPNet featuring cross-dimensional interaction, a novel encoder block with modified triplet attention, and atrous spatial pyramid pooling.
• Incorporated attention gating in modified skip connections to bridge the semantic gap between encoder and decoder.
• Utilized colonoscopy still images from Kvasir-SEG and CVC-ClinicDB datasets for training and validation.

Main Results:

• Achieved high performance with mean Intersection-Over-Union (mIoU) of 0.88 and Dice score of 0.9058 on Kvasir-SEG, and 0.8969 mIoU and 0.9372 Dice score on CVC-ClinicDB.
• Demonstrated generalizability on unseen datasets (Hyper-Kvasir, EndoTect 2020), establishing a new benchmark for polyp segmentation.
• Li-SegPNet excelled in segmenting medium-sized polyps, achieving 0.9086 mIoU and 0.9137 Dice score on Kvasir-SEG.

Conclusions:

• Li-SegPNet sets a new benchmark for polyp segmentation across multiple datasets.
• The model's lightweight design and high performance make it suitable for real-time clinical analysis in automated colorectal cancer diagnosis.