🚀 EvoGrayNet: Colon Polyp Detection and Segmentation using Evolutionary Network Architecture Search
🚀 Advanced detection and segmentation of colon polyps in endoscopy images
📌 Full code and pretrained models will be released soon!
Accurate colon polyp detection and segmentation in colonoscopy images remain challenging due to variability in appearance, size, and location. EvoGrayNet addresses this by integrating:
- Gray Module: Combines standard/depthwise separable convolutions with batch normalization and dropout to capture local-global features.
- Lightweight Attention Gate (AG): Refines feature maps for low-contrast regions and precise localization.
- Dilated Feature Extractor (DFE): Captures multiscale spatial context via dilated convolutions.
- Feature Recalibration (FR): Dynamically enhances channel-wise feature importance.
An evolutionary architecture search optimizes the model via crossover/mutation, maximizing Dice coefficient performance. EvoGrayNet outperforms 11 existing methods across 4 public datasets, achieving higher accuracy, lower FLOPs, and better generalization.
✅ High Accuracy: State-of-the-art Dice scores on polyp segmentation.
✅ Efficiency: Optimized FLOPs and parameters for clinical deployment.
✅ Robustness: Handles variability in polyp appearance, size, and lighting.
✅ Evolutionary NAS: Automated architecture search for optimal performance.
Stay tuned! ⏳