ARRS 2022 Abstracts


1980. Achieving Clinical-level Machine Learning System for Segmenting Organs on CT Scans
Authors * Denotes Presenting Author
  1. Jennifer Jin *; California State University, San Bernardino
  2. Soo Kim; Soongsil University
  3. Abner Wilding; Loma Linda University Medical Center
  4. Michael Christie; Loma Linda University Medical Center
  5. Shelley Villamor; Loma Linda University School of Medicine
  6. Abigail Beaven; Loma Linda University School of Medicine
  7. Daniel Jin; Loma Linda University Medical Center
Segmenting organs in medical images using deep learning is becoming a feasible approach to complementing organ segmentation by medical experts. Current approaches are limited to identifying target organs and these commercial tools show performance of ~75%. The objective of our research is to develop clinically applicable machine learning system for segmenting various solid organs in CT scans with improved segmentation performance than the current commercial standard.

Materials and Methods:
We trained Mask Region Based Convolutional Neural Network (Mask R-CNN) models to segment 4 organ types: liver, kidney, pancreas, and spleen. This retrospective IRB-approved study included 196 liver-protocoled CT series (66,918 CT slices), acquired from Loma Linda University Medical Center (LiTS from MICCAI17, and 3Dircadb); 300 kidney CT series (65,164 CT slices) were from KiTS19 and KiTS21 Challenges in MICCAI conferences; and 282 pancreas CT series (26,719 CT slices) and 41 spleen CT series (3,650 CT slices) from Medical Decathlon. The first software method set includes configuring optimal training sets, feature selection, hyper-parameter optimization, right-fitting with epochs, and performance tuning with measures. The second set includes software methods and machine learning models for post-processing initially segmented CT slices. These methods validated the continuity of organ appearance, consistency of organ location, size, shape, and Hounsfield units between contiguous CT slices. We trained a high performing Mask R-CNN model with configuration of optimal training sets. Then, we applied the post-processing to the segmentation results by applying five validation points.

We ran the organ segmentation by using popular commercial tools like 3D Slicer, ITK-SNAP, InVesalius, and TeraRecon. The performance was 47-81%, and the average performance was 74%. The variation is mainly due to the segmentation accuracy in the training set. The second experiment was to train a Mask R-CNN model without applying post-processing tactics. The performance range with our naïve tool was 54-80%, and the average performance was 75%. Our segmentation system without post-processing showed a similar range of performance to that of commercial tools. The third experiment was to apply the five post-processing tactics to the result of segmentation with the commercial tools and our initial tool. The performance range with the post-processing was 67-93%, and the average performance was 84%. The performance of our segmentation system with post-processing exceeded ~10% higher than the commercial tools and 9% higher than our naïve version.

We devised a set of five post-processing methods that can significantly improve the performance of organ segmentation in CT scans. An average of 84% segmentation performance of our system is not meant to be definitive, but subject to the sufficiency and quality of labels in the training set. Yet, our approach to validating the organ appearance continuity and organ image consistency between CT slices using post-processing was shown to be effective toward the clinical application.