A Comparative Study of Deep Learning Loss Functions: A Polyp Segmentation Case Study

Colorectal cancer is the third most common cancer diagnosed worldwide. The early detection of this disease can help in treating cases and save human lives. Deep learning algorithms appear as interesting tools used to successfully detect and segment polyps, thus improving surgical resection. When implementing these algorithms, there are different architectures to use. Depending on the architecture used, the obtained accuracy and the convergence speed of the algorithms can be different. To assess how these algorithms model datasets, we can use different metrics. In this paper, we compare different segmentation loss function metrics using one conventional architecture called U-Net. The performance was evaluated on three well-known polyp datasets, namely CVC-ClinicDB, Kvasir, and ETIS-Larib PolypDB. Findings show that the best results from CVC-ClinicDB are 85.89% for Dice, 84.89% from Kvasir, and 77.02% from ETIS-Larib PolypDB. The model behaves well even if we combine the three datasets. In fact, the accuracy level still reached 76.71%.