A Content-Driven Architecture for Medical Image Segmentation

As the U-Net architecture continues to push the state of the art in medical image segmentation, and keeping track of the multitude of derived models becomes increasingly difficult, opposing trends have emerged that prefer a coherent framework of ancillary processing tasks over the use of highly optimized and sophisticated models. This trend has culminated in the framework nnU-Net, which adapts preprocessing, training and inference to the respective dataset and thus managed to lead the Medical Segmentation Decathlon challenge despite relying on comparably simple models. In this paper, we focus on one ancillary technique that is commonly used but poorly addressed in literature: patchwise training. Since computational costs tend to increase drastically when using naive strategies, we discuss the benefits of content-sensitive sampling for patchwise training of deep learning segmentation models. We adapt this strategy in a content-driven architecture for abdominal aorta and stent-graft segmentation, where we evaluate and compare it with traditional sampling strategies based on a real-world clinical dataset.