Exploring Polyp Segmentation Through UNet-Based Models with Visual Insight

In view of the imperative role of Convolutional Neural Networks (CNNs) in medical imaging, particularly for polyp detection, this paper presents an advanced, efficient, and interpretable UNet-based model. This model is specifically designed for segmenting colorectal polyps, incorporating visual explanations to demystify the CNN's functioning. The study introduces modifications to the UNet encoder, incorporating different CNN architectures to strike a balance between computational efficiency and model complexity. The evaluation framework goes beyond mere accuracy, taking into account computational time and memory usage. By leveraging a robust dataset, the research emphasizes the significant impact of the encoder's type and depth on the accuracy of the UNet-based model. Among the models investigated, UNet-Densenet121, UNet-Densenet201, UNet-Resnet152 and UNet-EfficientNetb7 stand out, excelling in all evaluation criteria. Additionally, the study investigates the influence of image backgrounds and polyp attributes on model performance. In conclusion, the UNet-Densenet121 model stands out for its exemplary predictive capability and visual explanation, particularly in large-scale image contexts.