Architecture Overview

DFF-UNet targets skin-lesion segmentation with a compact U-shaped design that focuses on boundary precision and computational efficiency. The RFRE encoder refines feature extraction through residual connections and adaptive convolutional processing, while the PDCP bridge captures contextual lesion patterns using parallel dilated convolutions. The BSFD decoder then fuses semantic and spatial information through bottleneck-based skip fusion, allowing the model to preserve important lesion details without increasing complexity. This makes the architecture suitable for dermoscopic image analysis where precise segmentation and lightweight deployment are both important.

Architecture Diagram