An Analysis of Multi-organ Segmentation Performance of CNNs on Abdominal Organs with an Emphasis on Kidney

Mahmud Elahi Akhter, Ashfia Binte Habib, Rishad Arfin, FAHIMUL HAQUE, Syed Athar Bin Amir, ZUNAYEED BIN ZAHIR, SHAHRIAR HUSSAIN, RAJESH PALIT

Abstract: Manual hand tailored biomedical image segmentation of different organs is a time consuming and laborious task. However, for the last decade or so, many deep learning convolutional neural network (CNN) models have emerged claiming to have close to human level results on biomedical image segmentation of different type organs while automating the task. Multi-organ segmentation is the process of segmenting multiple organs of the same patient. This offers a convenient solution to automation by providing segmentation of multiple organs at a time. Since 2015, we seen massive improvements of deep CNNs. This has led to better multi-organ segmentation architectures and has influenced the study of multi-organ segmentation. In this paper, we analyze the performance of different multi-organ segmentation studies. Our main focus was kidney, spleen and pancreas. We emphasized on kidney with the believe that we will see an increase in kidney segmentation tasks and challenges. We found that multi-organ segmentation architectures have been improving over time and are performing quite well. However, we also found that there is substantial performance variance across the different studies even after using the same architecture and datasets on those studies.