Pyramidal Segmentation of Medical Images via Generative Adversarial Networks

Abstract

Colorectal cancer accounts for 10% of all cancer cases. Early detection is crucial for survival and is obtained by regular screening of the gastrointestinal tract for precursors of gastrointestinal cancer, known as polyps. Research has shown polyp miss rates of 14% to 30% for manual classification performed by doctors. Similar problems related to human error are observed when determining other attributes, such as borders and size of findings, which motivates the use of automated segmentation. Segmentation is the process of partitioning an image into areas with specified descriptions, meaning every pixel in the image is classified to detect and locate findings. In recent years, machine learning has provided impressive results for a wide variety of fields, ranging from language translation to facial recognition and cancer detection. The focus of this thesis will be to develop new segmentation models based on recent advances in machine learning and our hypothesis that learning several degrees of segmentation precisions by segmenting within grids may aid segmentation performance. This idea was motivated by the hypothesis that segmentation performance could be improved by building upon the knowledge of performing less precise segmentations. Our results suggest that segmentations of lower precisions produce better results at the cost of less precision, which proved useful for the cases where higher precision segmentations gave limited results. However, no impact on segmentation performance of higher segmentation precisions was observed. Generally, the normal pixel-level segmentation performance of our networks was as good as experiments with corresponding state-of-the-art neural networks for segmentation.