LGPS: A Lightweight GAN-Based Approach for Polyp Segmentation in Colonoscopy Images

Technical Summary: LGPS – A Lightweight GAN-Based Approach for Polyp Segmentation

Problem Statement
Colorectal cancer (CRC) remains a leading cause of cancer-related mortality globally, making early polyp detection and removal during colonoscopy critical for prevention. While deep learning-based segmentation methods have advanced significantly, they face three persistent limitations that hinder clinical adoption: (1) high computational costs that preclude real-time application; (2) difficulty in accurately segmenting small polyps or those with low contrast and blurry boundaries; and (3) poor generalization across diverse datasets due to variations in imaging conditions. Existing state-of-the-art (SOTA) models, such as U-Net variants, Transformers, and attention-based networks, often require substantial computational resources and struggle with these specific challenges.

Methodology
The authors propose LGPS (Lightweight GAN-based approach for Polyp Segmentation), a framework designed to balance high segmentation accuracy with computational efficiency. The architecture is built upon a Generative Adversarial Network (GAN) structure comprising a generator and a discriminator, enhanced with specific modules for medical image analysis:

1. Generator Architecture:

   • Backbone: The generator utilizes a modified MobileNetV2 backbone to ensure a lightweight design. Modifications include reducing the number of filters by a factor of two, removing the final classification layer and redundant intermediate layers, and employing depthwise separable convolutions with reduced expansion factors. This results in a model with only 1.07 million parameters.
   • ReSE Blocks: The network incorporates modified Residual Squeeze-and-Excitation (ReSE) blocks. These combine traditional residual connections with SE mechanisms to dynamically recalibrate feature maps, emphasizing diagnostically relevant regions while suppressing noise.
   • Decoder: The decoder employs an encoder-decoder structure with skip connections and progressive upsampling (bilinear interpolation) to recover spatial details, crucial for preserving the fine boundaries of polyps.

2. Discriminator Architecture:

   • The discriminator uses a patch-based adversarial framework to evaluate the quality of generated masks.
   • ConvCRF Refinement: A key innovation is the integration of Convolutional Conditional Random Fields (ConvCRF) layers within the discriminator. These layers apply learnable $3 \times 3$ convolutions followed by sigmoid activation to enforce spatial coherence and smoothness in the predicted masks while preserving edge details.

3. Hybrid Loss Function:

   • To address class imbalance (where polyp pixels are sparse compared to background) and ensure precise boundary definition, the generator is trained using a weighted combination of three loss functions:
     • Binary Cross-Entropy (BCE): For pixel-wise classification accuracy.
     • Weighted Intersection over Union (WIoU): To balance the contribution of foreground (polyp) and background regions.
     • Dice Loss: To optimize the overlap between predicted and ground truth masks, particularly effective for imbalanced classes.
   • The total loss is defined as $L_{total} = \lambda_1 L_{BCE} + \lambda_2 L_{IoU} + \lambda_3 L_{Dice}$, with weights set to 0.4, 0.3, and 0.3 respectively in the experiments.

Key Contributions

• Lightweight Design: LGPS achieves SOTA performance with only 1.07 million parameters, making it approximately 17 times smaller than the smallest existing SOTA model, facilitating deployment on resource-constrained devices.
• Architectural Innovations: The introduction of ReSE blocks for efficient feature extraction and ConvCRF layers for boundary refinement within a GAN framework specifically tailored for polyp segmentation.
• Robust Generalization: The framework demonstrates strong performance on unseen and challenging datasets, addressing the common issue of poor generalization in existing methods.

Experimental Results
The model was validated on five public benchmark datasets: Kvasir-SEG, CVC-ClinicDB, ETIS, CVC-300, and PolypGen.

• Performance on PolypGen: On the largest and most challenging test set (PolypGen), LGPS achieved a Dice coefficient of 0.7299 and an IoU of 0.7867, outperforming all compared SOTA methods (including U-Net++, ResU-Net++, and WDFF-Net).
• Performance on Other Datasets: On the CVC-ClinicDB dataset, the model achieved a Dice of 0.9261 and an IoU of 0.9238, surpassing several heavy-weight models. It also showed competitive results on ETIS and CVC-300.
• Ablation Studies: Experiments confirmed that the combination of BCE and Dice loss (BDice) was highly effective, though the full hybrid loss (3Loss) provided robust generalization. Removing ReSE or ConvCRF modules resulted in measurable performance drops, validating their necessity for feature extraction and boundary preservation.
• Qualitative Analysis: Visual comparisons indicated that LGPS better handles blurry boundaries, low contrast, and significant size variations (including small polyps) compared to existing methods, which often suffer from over-segmentation or missed detections in these scenarios.

Significance and Claims
The paper claims that LGPS represents a significant step forward in making automated polyp segmentation viable for real-time clinical applications. By drastically reducing model size without sacrificing accuracy, LGPS addresses the computational barriers that have limited the deployment of deep learning models in endoscopy suites. The authors emphasize that the model's ability to generalize well across diverse, unseen datasets (particularly PolypGen) underscores its potential reliability in real-world clinical settings where imaging conditions vary. The work positions LGPS as a new benchmark for efficient and accurate medical image analysis, offering a practical solution for improving early CRC diagnosis.