VGG-UNet Hybrid Architecture for Colorectal Polyp Segmentation: A Systematic Evaluation on CVC-ClinicDB

Accurate and automated detection of colorectal polyps in colonoscopy images is critical for early cancer prevention. While U-Net and its variants have demonstrated strong performance in medical image segmentation, the potential of combining pretrained convolutional encoders with the U-Net decoder architecture for polyp segmentation remains underexplored.

We propose a hybrid architecture that replaces the standard U-Net encoder with a pretrained VGG16 backbone, leveraging ImageNet transfer learning to improve feature extraction on the small CVC-ClinicDB dataset (612 images). We conduct a systematic ablation study evaluating the effects of loss function selection (BCE-Dice, Dice-only, Focal), encoder fine-tuning strategy (frozen vs. unfrozen), VGG backbone depth (VGG16 vs. VGG19), and data augmentation intensity (minimal, standard, heavy).

Our best configuration achieves a Dice coefficient of 0.9297 (validation) and 0.9156 (test), and an IoU of 0.8698 (validation), representing a significant improvement over the standard U-Net baseline (Val Dice: 0.7743, Test Dice: 0.7962). These results are competitive with recent state-of-the-art methods on CVC-ClinicDB, demonstrating that pretrained encoder transfer learning is an effective strategy for polyp segmentation on limited medical imaging datasets.