Semi-supervised Learning Methods for Semantic Segmentation of Polyps

Nowadays, colorectal cancer is one of the most common cancers, and early detection would greatly help improve patient survival. The current methods used by physicians to detect it are based on the visual detection of polyps in colonoscopy, a task that can be tackled by means of semantic segmentation methods. However, the amount of data necessary to train deep learning models for these problems is a barrier for their adoption. In this work, we study the application of different semi-supervised learning techniques to this problem when we have a small amount of annotated data. In this study, we have used the Kvasir-SEG data set, taking only 60 and 120 annotated images and studying the behaviour of the Data Distillation, Model Distillation, and Data & Model distillation methods in both cases, using 10 different architectures. The results show that as we increase the number of initially annotated data, most models obtained better results, but two of them performed worse in the baseline case. Furthermore, we can conclude that the Data Distillation method increases the performance of the models a 48.6% and 30.6% on average using 60 and 120 annotated images respectively. Finally, using only 12% of the annotated data and applying Data Distillation, the results obtained are not very far from those obtained by training the models with the fully annotated dataset. For all these reasons, we conclude that the Data Distillation method is a good tool in semantic segmentation problems when the number of initially annotated images is small.