A fiducial-less tracking method for radiation therapy of liver tumors by diaphragm disparity analysis part 1: simulation study using machine learning through artificial neural network

Abstract

Objective

The large respiratory-induced motion of the liver tumors can affect treatment planning and delivery in many ways. As a result, motion management techniques are necessary to mitigate these effects. An effective approach to reducing the effects of respiratory motion of liver tumors is real-time tracking of the tumor. The Cyberknife treatment modality uses a combination of kV X-ray images, LED markers, an optic camera, and surgically implanted fiducial markers to track liver tumors. However, the use of an invasive method for implanting fiducial markers can lead to complications. We propose a tracking method that requires no fiducial markers for liver tumors by using the projected location of the diaphragm to identify the 3D location of the liver tumor. With the use of the 4D extended cardiac-torso (XCAT) phantom, this simulation study aims to investigate the feasibility of localizing liver tumors through the tracking of the diaphragm-lung border.

Methods

An abdominal 4DCT dataset containing 20 phases of one breathing cycle was created by using the male model of the 4D XCAT phantom. One set of orthogonal DRR images (+ 45°) was generated for each phase. On each DRR image, an outline of the lung-diaphragm border was detected using an edge detection algorithm. The simulated tumor's gravity center was identified for each phase of the breathing cycle. Using artificial neural networks (ANNs), two respiratory scenarios correlating the diaphragm's location with the corresponding 3D location of the tumor were compared: (1) lung-defined tumor motion (TL) and (2) user-defined tumor motion (TA). Additionally, using the user-defined tumor motion, we also examined the accuracy of using ANN to track the tumor under the mismatched conditions during 4DCT reconstruction.

Results

Evaluation of the ANN model was quantified by the root mean square error (RMSE) values through the leave-one-out (LOO) validation technique. The RMSE for the TL motion was 0.67 mm and for the TA motion was 0.32 mm. When the ANN model was applied to the mismatched data, it generated the RMSE of 1.63 mm, whereas applied to the ground-truth data, the RMSE is 0.88 mm.

Conclusion

This simulation study shows that the diaphragm and tumor position are closely related. The developed diaphragm disparity-analysis approach, featuring tracking capability and verified with clinically acceptable errors, has the potential to replace fiducial markers for clinical application. The tracking method will be further investigated in clinical datasets from patients.

https://ift.tt/2mCZppG