An Investigation of AI Algorithms on Esophageal Gross Tumor Volume Segmentation.

Researchers

Journal

Modalities

Models

Abstract

The robustness and limitation of Deep Learning (DL) schemes for esophageal gross tumor volume (GTV) auto-segmentation has not been studied before on cancer stages and cancer position on the esophagus. Based on a large esophageal cancer dataset with accurate annotations from several institutions, we took the initiative to investigate this issue and try to find the underlying mechanism for unsatisfactory performance.We constructed a multi-center dataset containing CT scans of 238 patients from two hospitals with different CT scanners, with given information of cancer stages (II, III, IV) and cancer positions (up, middle, bottom) on the esophagus. We built the DL segmentation model on dataset from a hospital using a Philips scanner, and evaluate it on the other dataset, an external test set, from a hospital using a CMS scanner, using Dice Similarity Coefficient (DSC). Then we trained the model on different parts of the dataset (separated by cancer location on esophagus and stage) to investigate how cancer stage and position affect segmentation performance.Performance on validation and test sets from same hospital was similar (DSC of 77.1% and 76.9% respectively). Model trained on the dataset from a scanner did not perform as well on datasets from a CMS scanner (DSC of 63.2% for CMS), indicating that building DL models should consider a variety of training samples. Cancer stage was indeed most relevant to the segmentation performance. Late-stage cancer was comparatively much easier to predict for auto-delineation (69.2%, 81.3%, 83.8% on Dice score for stage II, III, IV, respectively). With regard to cancer position on esophagus, we observed that cancer located at the bottom part is most difficult to segment (67.1% on Dice score). When only computing the Dice score of slices containing cancer and ignoring those without cancer, it’s found that the score increases about 8.3% (77.1% to 85.4%). Deviated cancer localizing, instead of mislabeled deviated cancer boundary, may be the key reason for unsatisfactory segmentation.Early-stage gross tumor volumes located in the bottom part of esophagus are more difficult to give a concrete prediction. Models trained on data from bottom part outperform models trained on data from other positions. The learning schemes used here were capable of giving a preliminary delineation of esophageal gross tumor volume. This is significant for clinical doctors looking to reduce their workload and save medical resources.Copyright © 2021. Published by Elsevier Inc.