RSNA 2014 

Abstract Archives of the RSNA, 2014


INS169

Texture Analysis with Predictive Modeling of Solid Appearing Pancreatic Serous Cystadenomas versus Neuroendocrine Tumors

Scientific Posters

Presented on December 4, 2014
Presented as part of INS-THA: Informatics Thursday Poster Discussions

Participants

Franco Verde MD, Presenter: Nothing to Disclose
Siva P. Raman MD, Abstract Co-Author: Nothing to Disclose
Linda Chi Hang Chu MD, Abstract Co-Author: Nothing to Disclose
Yifei Chen BS, Abstract Co-Author: Nothing to Disclose
Peng Huang, Abstract Co-Author: Nothing to Disclose
Elliot K. Fishman MD, Abstract Co-Author: Research support, Siemens AG Advisory Board, Siemens AG Research support, General Electric Company Advisory Board, General Electric Company Co-founder, HipGraphics, Inc

PURPOSE

To investigate the ability of CT texture analysis (spatial frequency analysis) to distinguish between solid appearing, pancreatic serous cystadenomas (SCs) versus neuroendocrine tumors (NETs). Besides imaging appearance and clinical information, this tool hopefully allows for additional layer of confidence for differentiating between the two similar appearing tumors.

METHOD AND MATERIALS

IRB approval was obtained for retrospective review. Pathology proven 10 SCs and 10 NETs were obtained with similar CT appearance. Patients all had arterial and venous phase imaging. Both phases in 3 or 5 mm thick sections were uploaded to local server with TexRAD (TexRAD LTD, Somerset, UK). Region of Interest (ROI) was drawn around at all slices with discrete tumor visualized. Additional thresholding was applied to exclude attenuation values below -50HU. Analysis utilizes Laplacian of Gaussian spatial band-pass filters from 2mm to 6mm to highlight heterogenity of the ROI. Range of higher order statistics were obtained from the software (mean pixel intensity, entropy, standard deviation of pixel intensity, kurtosis, and skewness at each filter size. A random forest statistical model, already created at our institution, for other applications of TexRad on other tumors, was be applied to this dataset. This model will attempt to predict the histology in a prospective manner. Modeling was performed on 7 SCs and 7 NETs. Biostatistician, blinded to test patients, applied statistical model to each slice of test patients. A best guess to final diagnosis was made for each patient compared to radiologist leading diagnosis.  

RESULTS

Modeling: 360 arterial and 360 venous phase slices, from 7 SCs 240 arterial and 240 venous phase slices, from 7 NETs "Out-of-bag" error estimate rate: 5% Testing: 439 arterial and 439 venous phases from 4 SCs and 6 NETs. 4/4 SCs were correctly predicted. 4/6 NETs were correctly predicted Radiologist leading diagnosis: 0/4 SCs were correctly predicted 6/6 NETs were correctly predicted

CONCLUSION

Texture analysis has proven to be a superior adjunct to distinguish serous cystadenomas versus neuroendocrine tumors, difference between surgery and surveillance.

CLINICAL RELEVANCE/APPLICATION

Use of TexRAD (CT texture analysis) to predict histology of pathology proven, solid appearing, pancreatic serous cystadenoma versus neuroendocrine tumors.

Cite This Abstract

Verde, F, Raman, S, Chu, L, Chen, Y, Huang, P, Fishman, E, Texture Analysis with Predictive Modeling of Solid Appearing Pancreatic Serous Cystadenomas versus Neuroendocrine Tumors.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14002006.html