LCSegNet: Lightweight Context-Aware Segmentation Network

Overview

LCSegNet is a PyTorch-based semantic segmentation model designed for accurate tumor or object segmentation.
The architecture follows an encoder-decoder paradigm enhanced with spatial refinement and feature fusion mechanisms.

The model produces:

• Pixel-wise segmentation masks
• Boundary-aware predictions during training

Methodology

The network integrates:

• A ResNet34-based encoder for hierarchical feature extraction
• Spatial Information Enhancement (SIE) modules for refining spatial representations
• Feature Selective Fusion (FSF) blocks for multi-scale feature aggregation
• A decoder that progressively reconstructs high-resolution segmentation outputs

Boundary supervision is incorporated to improve edge localization and segmentation quality.

Installation

Clone the repository and install dependencies:

git clone https://github.com/Priyamvada28/LC-SegNet
cd Lcseg_net
pip install -r requirements.txt

Dataset

The model expects paired image-mask data:

• Input images: RGB
• Ground truth masks: binary (0 for background, 1 for target region)

Training

Run the training script:

python train.py

The training process optimizes a hybrid objective combining:

• Binary Cross Entropy loss
• Dice loss
• Boundary-aware loss

Model selection is based on validation loss.

Inference

Run prediction on trained weights:

python prediction.py

The model outputs binary segmentation masks obtained by thresholding predicted logits.

Results Interpretation

• Lower validation loss indicates better generalization
• A gap between training and validation loss may indicate overfitting

Conclusion

LCSegNet demonstrates an efficient approach to segmentation by combining spatial enhancement and feature fusion techniques within a lightweight architecture.