MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Automated Segmentation of Malignant Pleural Mesothelioma Tumor On Computed Tomography Scans Using Deep Convolutional Neural Networks

E Gudmundsson1*, C Straus2 , A Nowak3 , H Kindler4 , S Armato5 , (1) ,,,(2) University of Chicago, Chicago, Illinois, (3) University of Western Australia, Crawley, WA, (4) University of Chicago, Chicago, Illinois, (5) The University of Chicago, Chicago, IL

Presentations

(Tuesday, 7/31/2018) 4:30 PM - 6:00 PM

Room: Room 202

Purpose: Malignant pleural mesothelioma (MPM) tumor presents as pleural thickening on computed tomography (CT) scans. Volumetric segmentation of MPM has been a topic of interest in the measurement of MPM response to treatment. A deep convolutional neural network (CNN) was implemented for the segmentation of MPM tumor on CT scans.

Methods: Visible pleural thickening was contoured on the axial sections of 136 chest CT scans of 77 MPM patients using a semi-automated method. These contours were reviewed by a radiologist experienced in the diagnosis of MPM to establish reference segmentations. A deep CNN was trained on the two-class problem of differentiating between pleural thickening and normal thoracic tissue. A total of 13,489 axial sections containing segmented tumor from 122 of the contoured chest CT scans was used to train the network. A validation set of 1077 axial sections from the CT scans of six distinct patients not included in the training set was used to evaluate network performance during training. A test set of 736 axial sections from CT scans of ten distinct patients not included in the training or validation sets was used to evaluate the performance of the network after training by calculating the Dice Similarity Coefficient (DSC) between computer-generated and reference segmentations.

Results: Median DSC on the validation set was 0.84 (mean 0.77; standard deviation 0.21) after 12 epochs of training, after which validation performance diminished. Median DSC on the test set was 0.75 (mean 0.68; standard deviation 0.21) after 12 epochs of training.

Conclusion: A deep CNN was implemented for the first time for the task of automated segmentation of MPM tumor on CT scans. Future work will focus on improving the accuracy and robustness of this method by training on a larger set of annotated CT scans and exploring the use of three-dimensional convolutional filters.

Funding Support, Disclosures, and Conflict of Interest: Partially funding was provided by the NIH S10 RR021039 and P30 CA14599 grants, and by the Plooy Family and the Kazan McClain Partners' Foundation, Inc.. SGA receives royalties and licensing fees for computer-aided diagnostic technology through the University of Chicago and is a consultant for Aduro Biotech, Inc..

Keywords

Segmentation, CT, Quantitative Imaging

Taxonomy

IM- Dataset analysis/biomathematics: Machine learning

Contact Email