Siamese-Based Deep Learning for Markerless Lung Tumor Tracking During Stereotactic Radiotherapy.

Purpose/Objective(s) During stereotactic lung radiotherapy we want to monitor the target position to ensure it remains inside the planning target volume (PTV). This ensures accurate treatment delivery and may allow for margin reduction. Many LINAC systems have a kilo-voltage (kV) imaging system mounted orthogonally to the MV beam. This can acquire images during volumetric modulated arc therapy delivery. The 2D-kV images can be used to identify the tumor position, e.g., using template matching and triangulation (TMT). However, small lung targets are difficult to track, due to e.g., low tumor contrast, occlusion (e.g., by spine), or noisy images. We investigated deep learning (DL) as an alternative method for lung tumor tracking. Materials/Methods Clinical data from 6 patients (P1-P6) with small lung tumors (0.5-3 cm diameter) and a 3D-printed anthropomorphic phantom (P0), including a low-density lung tumor, was used. For each patient, data included 4D-CT scans, fluoroscopic kV images and breathing-related surface motion (Real-Time Position Management, RPM). For the phantom, a 3D-CT scan and kV images taken with a static and moving couch (regular and irregular motion) were available. Due to lack of ground truth tumor locations for the kV images, a patient-specific DL model is trained on Digitally Reconstructed Radiographs (DRRs), generated from the planning CT scans. To enhance the tumor's visual diversity in DRRs, we augmented the training data with locally deformed CT scans. A Siamese-based tracking by similarity method is hypothesized to be better suited for the tumor tracking task than other DL methods, due to its more localized search space and the use of previous frames (temporal information). Results Table 1 shows the superior-inferior (SI) direction results obtained by the Siamese model and TMT, on all kV images. For the patient data, the Correlation coefficient between estimated tumor trajectory and RPM was computed, due to lack of ground truth locations. For the phantom, the Mean Absolute Distance (MAD) between detected and ground truth center coordinates was computed. For all 6 patients and the phantom, the Siamese model outperformed TMT. Conclusion A novel approach for real-time markerless lung tumor tracking is presented, based on deep Siamese networks. The model compares favorably to the template matching/triangulation-based technique and merits further investigation. The current analysis presents only SI tracking. Reliable 3D tracking, with sub-mm accuracy, is the ultimate goal. Further work is needed to achieve this.