Optimized SegNet for Gastric Polyp Segmentation on Colonoscopy Images

Gastric polyp segmentation is essential in helping clinicians detect and remove polyps to prevent their potential progression to cancer. However, manually segmenting polyps is a laborious task and demands costly human resources. To overcome this issue, deep learning approaches such as SegNet have been applied for automated polyp segmentation. Though SegNet has proven to obtain good results, its restricted receptive field may not capture extensive contextual information from input data, thereby reducing its robustness. This study proposes an optimized SegNet model that incorporates dilation rate and dropout within the convolution layers to increase its receptive field and prevent overfitting during the model training. The model is trained, validated, and tested on colonoscopy images in the Kvasir-SEG dataset. The optimized SegNet model is cross-validated to evaluate its performance. The proposed model is evaluated by comparing it with the other models, and the results show that the optimized SegNet achieves 99.81% of accuracy, 99.24% of sensitivity, 99.93% of specificity, 99.23% of F1-score, and 0.99 of mIoU. Conclusively, the optimized SegNet model proves its superiority in segmenting gastric polyps compared to existing state-of-the-art models.