Sammendrag
Colorectal cancercolorectal cancer (CRC) is a widespread disease which is a threat to public health. Abnormalities in the colon, like polyps, can become cancerous. It is important to detect polyps early, in order to prevent a potential spread of cancer. Polyps can be overlooked during screening, and typically doctors have a polyp miss rate ranging from 14 to 30%. Several promising computer systems have been developed to help doctors lower their polyp miss rate. Obtaining a large, high quality dataset is important when building such a system, and the lack of data is perhaps the biggest challenge in the field today. Data is arguably the most valuable resource in machine learning. Complex neural networks are dependent on great amounts of data in order to perform well. The colon is full of complicated structures, and a dataset should contain examples of as many examples of both healthy and unhealthy structures as possible. However, medical data is hard to get hold of due to legal restrictions and the cost of performing examinations. Currently a highly qualified, medical expert is needed to annotate data as well, further complicating matters. We have developed a system which can take an existing dataset and use it to generate new, artificial data which can be added to the dataset. This will make it easier to create a large enough dataset for polyp detection systems. In other words, we can generate real-looking videos of polyps. A total of 41 generated videos was provided to two medical experts, and they were asked to comment on the quality of the videos. Their comments revealed that shapes and colors in the videos look real. They additionally stated that they found these videos relevant for detecting other abnormalities in the colon. We also trained two polyp classifiers on the same dataset, but for one of the classifiers we also added our artificial videos. We found that the results were inconclusive, though we believe that it should be possible for the artificial videos to improve performance.