(Monday, 7/30/2018) 1:15 PM - 1:45 PM
Room: Exhibit Hall | Forum 5
Purpose: To investigate localizing liver tumors with artificial neural networks through tracking lung-diaphragm border
Methods: In this study, an abdominal 4DCT dataset containing one breathing cycle (8 or 20 phases) was previously acquired for three patients. One set of DRR images (+45 or -45 degrees) was generated for each phase. On each DRR, an outline of the lung-diaphragm border was detected using an edge detection algorithm. The tracking volume’s gravity center was identified for each phase of the breathing cycle using a MATLAB program, and this serves as the measured center of that volume. To correlate the diaphragm’s 2D location with the corresponding 3D location of the tracking volume, a model was generated using artificial neural networks (ANN). Five models were generated for each of the 5 target volumes. The testing of these models was done by comparing the model-predicted tracking volume location to the measured tracking volume location using the mean root squared error (MRSE) values and the leave-one-out (LOO) validation technique.
Results: For the best model, the mean difference between the model-predicted and the measured tracking volume location was 0.032mm, 0.003mm and 0.008mm in the anterior-posterior(AP), lateral and superior-inferior(SI) directions respectively. For the worst model, the mean difference between the model-predicted and the measured tracking volume location is 0.13 mm, 0.13mm and 0.12mm in the AP, lateral and SI directions respectively. Over the 5 models, the average MRSE was 1.77 ± 0.60mm with a range of 1.03mm to 2.80mm.
Conclusion: Gold fiducial markers, requiring surgical procedure to be deployed, are conventionally used in radiation therapy. This study showed that the diaphragm and tracking volumes are closely related. The developed approach, verified with clinically accepted errors, has the potential to replace fiducial markers for clinical application. The tracking method will be further investigated in a larger cohort of patients.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
