An adjacent-frame Siamese U-Net for markerless angle-agnostic lung tumor tracking from projections in radiotherapy.

BACKGROUND: Respiratory motion and inter-fraction anatomical variations reduce the accuracy of dose delivery in lung cancer radiotherapy. Projection-based tumor tracking methods are limited by anatomical overlap and weak tumor contrast. This study developed a deep learning-based real-time, angle-agnostic tumor tracking method for predicting tumor contours from projections. METHODS: Three 4DCTs were collected for each of eight lung cancer patients. Deformable registration was performed by registering the 50% respiratory phase to the remaining phases using Elastix to generate deformation vector fields, which were used to generate intermediate motion states in the first 4DCT. Localized Gaussian deformation was applied to simulate tumor morphology variations. Projections and tumor masks were generated, and the contrastive unpaired translation model was applied to the projections. A Siamese U-Net incorporating a shared-weight encoder and a multi-modal change fusion module was designed for angle-agnostic tumor tracking using two adjacent projections, referred to as SiamMCF-UNet, and was compared with models reported in previous studies. After patient-specific training using the first 4DCT, tracking performance was evaluated using the Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95) across projection angles and respiratory phases on the second and third 4DCTs. RESULTS: SiamMCF-UNet achieved significantly higher DSC than both comparison models and lower HD95 than U-Net in both evaluations. Mean DSCs were 0.937 ± 0.013 and 0.914 ± 0.040, and mean HD95 values were 3.23 ± 0.96 mm and 5.45 ± 4.83 mm. Performance was generally consistent across respiratory phases, although angle-dependent variations were observed across gantry angles. CONCLUSION: SiamMCF-UNet demonstrated accurate and robust real-time, angle-agnostic lung tumor tracking.