FA-DGNet: Pseudo-Supervised Difficulty-Aware Learning and Guided Refinement for Colonoscopic Polyp Segmentation

Colonoscopy is a crucial method for colorectal cancer screening and the early detection of precancerous polyps; however, accurate polyp segmentation is challenged by factors such as indistinct boundaries, minute or flat structures, and regions of high-intensity interference. Given the spatial variability of these challenges, we propose FA-DGNet, a polyp-segmentation model built upon the HarDNet-68 backbone that integrates “pseudo-supervised difficulty learning” with “guided refinement” strategies. The model employs a deterministic training objective that incorporates dilated mask boundaries, locally sparse foreground regions, and brightness-inferred areas prone to artifacts, all without requiring additional manual annotations. The learned difficulty response modulates two stages of the decoder via a bounded residual amplification mechanism, while auxiliary supervision is provided by dedicated artifact and edge detection heads. On the Kvasir-SEG dataset, FA-DGNet achieved a mean Dice coefficient of 0.9071 and a mean Intersection over Union (IoU) of 0.8508, outperforming the baseline model’s metrics of 0.9006 and 0.8418, respectively. Furthermore, cross-dataset transfer experiments and evaluations across five independent test sets confirmed the model’s generalization capability and performance stability. Static performance analysis indicates that the model incurs only marginal increases of approximately 0.43% in active parameters and 0.45% in multiply–accumulate (MAC) operations. Overall, FA-DGNet offers a low-overhead, difficulty-aware optimization strategy, providing a practical solution for colonoscopic polyp segmentation.