DCD-DeepLabv3+: Integrating Attention and Dense Feature Fusion for Colorectal Polyp Segmentation
TL;DR
Results show that the proposed DCD-DeepLabv3+-based model improves the DeepLabv3+ baseline on the evaluated polyp-segmentation benchmarks, and further validation is required before clinical use.
Abstract
Accurate segmentation of colorectal polyps can support computer-assisted analysis of colonoscopic images, but variation in polyp morphology, low lesion-to-background contrast, and indistinct boundaries remain challenging. We propose DCD-DeepLabv3+, an enhanced DeepLabv3+-based model designed to improve polyp segmentation across diverse colonoscopy datasets. DCD-DeepLabv3+ extends a MobileNetV2-based DeepLabv3+ backbone with a dual-channel attention (DCA) module, a convolutional block attention module (CBAM), and a dense block (DB). DCA reweights feature channels, CBAM refines channel-wise and spatial responses, and DB promotes feature reuse through dense connections. A multilevel feature-fusion strategy further combines low-level spatial detail with high-level semantic information. We evaluated the model on five public datasets. On Kvasir, DCD-DeepLabv3+ achieved 91.70% Dice and 87.37% intersection over union (IoU), exceeding the DeepLabv3+ baseline by 0.97 and 1.58 percentage points, respectively. Accuracy, precision, and recall increased by 1.55, 2.40, and 1.01 percentage points. The model also achieved the highest IoU among the compared methods on ClinicDB, CVC-300, ColonDB, and ETIS. Ablation results supported the combined contribution of DCA, CBAM, and DB. These results show that the proposed feature-enhancement strategy improves the DeepLabv3+ baseline on the evaluated polyp-segmentation benchmarks. Further validation is required before clinical use.