P1.06-04 Deep Learning-Based Segmentation of Mesothelioma on CT Scans: Application to Patient Scans Exhibiting Pleural Effusion

Abstract

Tumor volume has been a topic of interest in the evaluation of treatment response, staging and prognosis of malignant pleural mesothelioma patients. Many mesothelioma patients present with or develop pleural fluid, the presence of which may complicate tumor segmentation on CT scans. We implemented a method for the automated segmentation of mesothelioma tumor on CT that explicitly excludes pleural effusion, which, if included in the segmentation of tumor, could confound the assessment of tumor volume. Deep convolutional neural networks (CNNs) were trained for the segmentation of mesothelioma tumor in each hemithorax. A database was collected of 180 CT scans of 160 mesothelioma patients who exhibited tumor and pleural effusion. 6026 axial sections containing segmented tumor (1243 sections exhibiting pleural effusion) from 134 chest CT scans were used to train deep CNNs for segmentation of mesothelioma tumor. A radiologist contoured tumor on a test set of 94 axial sections that exhibited both tumor and pleural effusion; these sections were randomly selected from 46 CT scans of 34 patients not included in the training set. Performance was evaluated on the test set by calculating the Dice Similarity Coefficient (DSC) between computer-generated and reference segmentations; DSC is a measure of overlap between a pair of segmentations (a value of 0 indicating no overlap, 1 indicating complete overlap). We compared the performance of the present method to a previously published deep learning-based method for the automated segmentation of mesothelioma tumor on CT scans; differences in DSC values achieved on the test set by the two methods were assessed through a two-tailed paired Wilcoxon signed-rank test. A boxplot of DSC values achieved on the test set by the current method and the previously published method is shown in Fig. 1. The median DSC on the test set achieved by the current method was 0.66 (inter-quartile range 0.20); the median DSC on the test set achieved by the previously published method was 0.51 (inter-quartile range 0.34). The difference in DSC between the two methods was statistically significant (p < 0.0001). A deep CNN was implemented for the task of automated segmentation of mesothelioma tumor on CT scans of patients who also exhibit pleural effusion. The present method achieved a statistically higher overlap (p <0.0001) with radiologist-provided reference contours than a previously published method on a test set of 94 axial CT sections of mesothelioma patients exhibiting both tumor and pleural effusion.