A Lightweight DeepLabv3+ Model for Real-Time Polyp Segmentation in Colonoscopy Images

Abstract

Automatic and accurate segmentation of polyps in colonoscopy images is considered crucial for the early detection of colorectal cancer, which helps improve patient survival rates. However, achieving high segmentation accuracy while maintaining low computational complexity is recognized as a major challenge, particularly for real-time clinical applications. In this paper, a lightweight and optimized DeepLabv3+ architecture is proposed, specifically tailored for efficient and precise polyp segmentation. The original DeepLabv3+ encoder is replaced with a MobileNetV2 backbone to enhance computational efficiency by significantly reducing the number of parameters, while maintaining segmentation accuracy. Additionally, a Channel Attention mechanism is integrated to recalibrate feature responses, selectively emphasizing informative features while suppressing irrelevant ones. Atrous Spatial Pyramid Pooling (ASPP) is employed to capture multi-scale contextual information without reducing spatial resolution. Furthermore, the decoder module is designed to combine features at multiple levels and refine them progressively, accurately reconstructing segmentation masks with clear boundaries. The proposed model is evaluated on the CVC-ClinicDB dataset, achieving a Dice coefficient of 0.9201 and an Intersection over Union (IoU) of 0.8547 while maintaining low computational overhead. These results demonstrate the effectiveness and potential of the model for real-time clinical deployment in resource-constrained environments.