SCA-Net: A Scale- and Contrast-Aware Network for Subtle and Low-Contrast Polyp Segmentation

This repository contains the PyTorch implementation of SCA-Net, a scale- and contrast-aware framework for robust polyp segmentation under substantial scale variation and low-contrast boundary conditions.

Method

SCA-Net follows an encoder-decoder design and integrates three complementary components:

• Semantic Module Group (SMG), consisting of a Cross-Scale Global Aggregator (CSGA) and Gated Semantic Injection (GSI), builds a shared cross-scale semantic token space and injects aggregated context into stage-wise features through gated residual modulation.
• Size-Adaptive Dynamic Router (SADR) introduces a scale-supervised soft routing mechanism that uses mask-derived size labels to adaptively balance receptive-field experts according to object scale.
• Laplacian-Guided Synergistic Refiner (LGSR) structurally couples reverse semantic guidance, Laplacian-guided edge modulation, and local-context fusion to improve boundary localization for weakly demarcated lesions.

Through this coordinated design, SCA-Net aims to produce more discriminative, scale-adaptive, and boundary-sensitive representations for subtle and low-contrast polyp segmentation scenarios.

Results

Experiments are conducted on five public polyp segmentation datasets: Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB. Following the common protocol, 900 images from Kvasir-SEG and 550 images from CVC-ClinicDB are used for training, while CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB are used as unseen datasets for cross-dataset generalization.

Dataset	Setting	Backbone	mDice (%)	mIoU (%)
CVC-ClinicDB	Seen	ConvNeXt-T	94.7	90.3
Kvasir-SEG	Seen	ConvNeXt-T	93.3	88.7
CVC-300	Unseen	ConvNeXt-B	92.1	86.2
CVC-ColonDB	Unseen	PVTv2-B4	81.2	73.4
ETIS-LaribPolypDB	Unseen	PVTv2-B4	86.0	78.7

On the challenging unseen ETIS-LaribPolypDB dataset, SCA-Net with a PVTv2-B4 backbone achieves 86.0% mDice and 78.7% mIoU.

The qualitative results cover challenging cases involving irregular boundaries, small or elongated structures, subtle appearance, and extremely small targets. Compared with competing methods, SCA-Net yields more complete foreground prediction, clearer boundary continuity, and fewer false activations in challenging regions.

Installation

Recommended environment:

• Python 3.10.15
• PyTorch 2.10.0
• CUDA 12.8

Install dependencies:

pip install -r requirements.txt

Install the CUDA-enabled PyTorch build that matches your local CUDA toolkit or driver.

Data Preparation

The dataset settings are consistent with the common polyp segmentation protocol. Place the datasets under data/ with the following structure:

data/
  TrainDataset/
    images/
    masks/
  TestDataset/
    CVC-300/
      images/
      masks/
    CVC-ClinicDB/
      images/
      masks/
    Kvasir/
      images/
      masks/
    CVC-ColonDB/
      images/
      masks/
    ETIS-LaribPolypDB/
      images/
      masks/

Training and Evaluation

Train SCA-Net:

python train.py

Run prediction:

python predict.py --checkpoint ./checkpoints/sca_net/epoch_100.pth

Evaluate predictions:

python evaluate.py --pred-root ./results/sca_net

Checkpoints

The checkpoint will be released after the paper is published.