Rapid segmentation of thoracic organs using u-net architecture

Medical imaging provides a non-invasive method to diagnose, monitor, and plan the treatment of disease inside the human body. The increasing prevalence of radiological scanners and prescription of their use has presented a significant challenge for radiologists in accurately diagnosing disease whilst dealing with a growing number of scans to review. Recent advances in Artificial Intelligence (AI), especially in machine learning, are enabling researchers to improve the patient experience, enhance the planning of medical treatments and increase the rate of examination of scans. In this study, a 2-dimensional (2D) U-net based deep learning model was used to automatically segment five organs of interest from Computed Tomography (CT) scans of the thoracic region. Comparable results were achieved in comparison to the top seven models from a prior thoracic organ segmentation challenge. The framework can perform the segmentation tasks within 20 seconds, reducing workload for radiologists and increasing throughput. This study shows that a simple U-net based framework can be sufficient for the task at hand rather than pursuing much more complicated architectures, depending upon the complexity of the problem. Furthermore, we investigated the effect of 3D interpolation on dice scores in anticipation of further research applications in mapping segments to a 3D volume render. We find performance degradation with respect to the dice score after mapping the masks to original dimensions.