| Popis: |
Automatic polyp detection systems are an important tools to aid in the diagnosis and prevention of colorectal cancer. Currently, methods based on deep learning approaches have presented promising results. However, the performance of these techniques is highly associated with the use of large and varied data samples for training. This is one of the main limitations of applying Deep Learning techniques in the medical field since the amount of data for training is generally limited compared to nonmedical disciplines. This work proposes a novel method to increase the quantity and variability of training images from a publicly available colonoscopy dataset. The developed approach enrich the training data adding polyps to regions of nonpolypoid samples, creating automatically new data with their appropriate labels. Performance results show that convolutional neural networks trained in these syntactically-enhanced datasets improved the accuracy on polyps segmentation task while reducing the false positive rate. These results open new possibilities for advancing the study and implementation of new methods to automatically increase the number of samples in datasets for computer-assisted medical image analysis. |
