LGPS: A lightweight GAN-based approach for polyp segmentation in colonoscopy images

Accurate and real-time polyp segmentation is essential for early colorectal cancer detection, yet remains challenging due to small, low-contrast lesions and the computational demands of many deep learning models. We propose LGPS, a Lightweight GAN-based Polyp Segmentation framework that achieves high segmentation accuracy with minimal computational overhead. LGPS integrates a MobileNetV2-based generator enhanced with Residual Squeeze-and-Excitation (ReSE) blocks and a discriminator equipped with Convolutional Conditional Random Fields (ConvCRF) to refine spatial coherence and boundary precision. A hybrid loss function — combining Binary Cross-Entropy, Weighted IoU, and Dice Loss — improves class imbalance handling and sensitivity to small or blurry polyps. LGPS is an extremely compact model, requiring only 1.07 million parameters — over 17× smaller than many recent transformer- and CNN-based SOTA architectures — while preserving high segmentation accuracy. In quantitative evaluation, LGPS achieves a Dice score of 0.7299 and an IoU of 0.7867 on the multi-center PolypGen dataset, the most challenging benchmark in polyp segmentation. On the widely used CVC-ClinicDB dataset, LGPS attains the highest IoU (0.9238) among lightweight and transformer-based approaches. The model also runs at 100.08 FPS on 256 × 256 inputs, demonstrating true real-time capability. Stratified evaluation and qualitative results further confirm its robustness on small and low-contrast lesions. These results highlight LGPS as a computationally efficient yet high-performing framework suitable for real-time clinical deployment. The code is available at https://github.com/Falmi/LGPS/.