Radiation and Cancer Physics
SS 15 - Physics 3 - Treatment Planning
109 - Automated Contouring of Contrast and Non-Contrast CT Liver Images With Fully Convolutional Neural Networks
Monday, October 22
4:35 PM - 4:45 PM
Location: Room 006
Automated Contouring of Contrast and Non-Contrast CT Liver Images With Fully Convolutional Neural Networks
B. M. Anderson1, E. Lin1, C. E. Cardenas1, E. J. Koay2, B. Odisio1, and K. K. Brock2; 1The University of Texas MD Anderson Cancer Center, Houston, TX, 2University of Texas MD Anderson Cancer Center, Houston, TX
Purpose/Objective(s): Liver contouring can be particularly labor-intensive based on size and unique patient liver shape. While model-based segmentation methods have shown promise in the past, they can be relatively slow (> 10 minutes), and not robust to unique patient anatomy. Fully-convolutional neural networks have shown great promise in the ability to accurately segment 2D images with multiple classes in short time periods and robustness against nuisance variations. We hypothesize that fully-convolutional networks (FCNs) can be used to rapidly and to accurately contour the
Materials/Methods: Manual contours of the liver were delineated on the CT scans of 58 patients who had colorectal liver metastases (CLM) by a trained radiologist. A previously trained FCN, the Visual Geometry Group (VGG)-16 network, was adapted for rapid segmentation of the liver. A skip-layer architecture, similar to that found in the popular ‘U-net’ design, was implemented in order to improve output segmentation results. Various batch sizes, learning rates, and filters were investigated before deciding on the final model. For validation, 161 contrast-enhanced image sets released by the Medical Imaging Computing and Computer Assisted Intervention (MICCAI) challenges were obtained, as well as 10 non-contrast-enhanced image sets from the institution. The contrast-enhanced image sets came specifically from the abdominal challenge (n=30) and liver tumor segmentation challenge (LiTs, n=131). As a final test, 26 patients with clinically defined liver contours were compared by a blinded diagnostic radiologist. Upon the presentation of the two contours, responses were categorized based on the contour preference, and clinical usability.
Results: The final model was selected based on its performance in dice similarity coefficient (DSC) scores. The average DSC and standard deviations for the validation groups were as follows: µAbdomen:0.93, σAbdomen:0.02, µLiTS:0.92, σLiTS:0.05, and µNon-Contrast:0.95, σNon-Contrast:0.006. Of the contours generated from this model compared to 26 clinically defined contours, 50% (13/26) were preferred to the previously created contour, 54% (14/26) would be clinically useful without edits, 92% (24/26) useable or useable with minor edits (<1 minute of work), and 8% (2/26) required major edits. Liver contours were predicted, post-processed, and transferred to the planning system within 1 minute using an Intel i5 processor (3.3 GHz) and 16 GB RAM.
Conclusion: These results suggest that the developed FCN can rapidly and accurately contour the liver on CT scans in a manner similar to those contoured in our clinic. Beyond the scope of DSC, on a small subset of cases these contours have been shown to be clinically useful and even preferable to currently created manual contours, warranting further investigation. For the two case of the major edits, these patients had extensive disease which was not seen in the training data, and will be investigated further in future iterations of the model.
Author Disclosure: B.M. Anderson: None. E. Lin: None. E.J. Koay: Research Grant; MD Anderson Cancer Center, PANCAN-AACR, Philips, Radiological Society of North America. Patent/License Fees/Copyright; Pending Patent. B. Odisio: None. K.K. Brock: Research Grant; RaySearch Laboratories. Honoraria; RaySearch Laboratories. Royalty; RaySearch Laboratories.