Evaluation of the reliability of markerless tumor tracking with single-energy and dual-energy imaging using machine...

Markerless tumor tracking (MTT) using single-energy (SE) kilovoltage (kV) imaging has been proposed as a technique for lung tumor motion management. However, bony structures can obscure the tumor and make tracking challenging. Dual-energy (DE) subtraction imaging can suppress the bone and therefore enhance tumor visibility, potentially improving tracking accuracy.

To develop a machine learning model that predicts the reliability of MTT for SE and DE imaging using phantom and patient data.

Images were collected at high and low energy (120/60 kV) using fast-kV switching on dynamic thorax phantom with 5-15 diameter targets and 20 lung cancer patients, including one patient with two tumor locations. DE images were then generated off-line using weighted logarithmic subtraction. Separately, a template-based tracking algorithm was used to track the tumor on both phantom and patient images. For phantom data, the ground truth (GT) was defined by the cos⁴ waveform programmed into the motion controller. Since the GT was unknown for patient data, it was estimated using a Kalman filter and further refined using visual inspection. The tracking success rate (TSR) was subsequently calculated. Based on TSR, patients were stratified into three tracking groups: good, moderate, and poor. A logistic regression (LR) model was implemented with tracking features, including match score, peak to side-peak ratio (PSR), x- and y-velocity. To address the distribution differences between phantom and patient data, a domain indicator and interaction features were also incorporated. The model was trained on phantom and 17 patient data using leave-one-out cross-validation, and its performance was evaluated on three held-out patients, one from each tracking group. Separate LR models were developed for each imaging modality, resulting in SE-LR and DE-LR models. For each held-out patient, the confusion matrix, sensitivity, and specificity were reported.

For both SE-LR and DE-LR models, PSR has the largest coefficient value, indicating that it is the most influential factor when determining the reliability of tracking. At a specificity of 95%, the DE-LR model achieved significantly higher sensitivity than the SE-LR model (0.908 vs. 0.731, p < 0.01 for phantom data and 0.657 vs. 0.589, p < 0.01 for patient data). In the three held-out patients, the DE-LR model outperformed the SE-LR model across all metrics. The largest gains were found in the patient with good tracking quality.

The proposed LR models successfully predicted the reliability of SE- and DE-MTT. Across phantom and patient datasets, the DE-LR model consistently demonstrated improved performance compared with the SE-LR model, suggesting that DE imaging may be beneficial for MTT applications.