MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Relabeling Non-Standard to Standard Structure Names Using Geometric and Radiomic Information

W Sleeman1*, J Palta1, P Ghosh2, R Kapoor1, (1) Department of Radiation Oncology, Virginia Commonwealth University, Richmond, VA (2) Department of Computer Science, Virginia Commonwealth University, Richmond, VA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
To harness the potential of Big Data in Radiation Oncology, there has been growing efforts to prospectively utilize the TG-263 compliant structure names for targets and OARS. These efforts do not address the retrospective data or clinics where these recommendations have not yet been adopted. There have been attempts in the past to automatically relabel the structure names to TG-263 compliant names using the traditional geometric and anatomical features. In our work, we propose to improve the accuracy of these methods by including common radiomic features.

Methods:
We used DICOM data from 614 lung and 654 prostate patients across 39 treatment centers which resulted in 11,219 and 12,474 total structures respectively. Structures chosen to be automatically labeled were: (lung) esophagus, heart, PTV, spinal cord, brachial plexus; (prostate) rectum, bladder, right femur, left femur, PTV, small bowel, large bowel; all other present structures formed a non-annotated structure group. These structures were converted to volumetric bitmap representations and were then reduced to 100 features using singular value decomposition. Radiomic features such as median, skew, kurtosis, signal-to-noise ratio, and uniformity were also calculated for each structure using the underlying soft tissue and dose information. From the Apache Spark platform, the Random Forest classifier with 5-fold cross-validation was used to compare the accuracy of predicting the correct TG-263 structure name with geometric information with and without radiomic features.

Results:
Including radiomic features provided F1-score changes of -0.002 to 0.215 with an average improvement of 0.053 across all structure types. The the lung PTV and rectum F1-scores were most affected with improvements of 0.215 and 0.106 respectively. The most common misclassification was between OARS and the non-annotated structures.

Conclusion:
This works shows that soft tissue and dose related radiomic features can be used to improve the predictive power of a structure name relabeling system.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email