End-to-End Automation Enables High Throughput Spleen Volume Assessment for Computed Tomography Clinical Trials

Abstract

Delineation of CT abdominal anatomical structure, especially spleen segmentation is essential for measuring tissue volume and biomarkers, so that it can be utilized not only as liver diseases and infection diagnosis purposes, but also as computer-based surgery planning. Recently, an increasing amount of CT data hampers clinicians from manual segmenting. So segmentation algorithm using deep learning is robustly implemented to solve this issue. However, computerized segmentation also has had difficulties (1) managing intermediate results (e.g., resampled scans, 2D sliced image for deep learning), and (2) setting up the system environments and packages. To treat this challenge, we propose an automated pipeline for abdominal spleen segmentation. This pipeline provides end-to-end containerized process that allows user not to install any packages and not have to deal with the intermediate results locally. The pipeline has three major stages which are preprocessing of input data, segmentation spleen using deep learning, reconstructing 3D with generated label, and merger into a pdf by matching with the original image dimension and for later demonstration purposes. Regularly, DICOM subject that has 150 scans takes 30 minutes in average for only manual segmentation while Proposed pipeline only spends 50 seconds for 150 scan DICOM subject. Even if it includes all the preprocessing and setups, once the container localized, the whole process elapses 20 minutes from beginning to end.