Author: Weixing Cai, Laura I. Cervino, Qiyong Fan, Yabo Fu, Tianfang Li, Xiang Li, Jean M. Moran, Hai Pham, Pengpeng Zhang 👨🔬
Affiliation: Department of Medical Physics, Memorial Sloan Kettering Cancer Center, Memorial Sloan Kettering Cancer Center 🌍
Purpose: AAPM Task Group Report 273 emphasizes the importance of rigorous validation to ensure the generalizability and robustness of machine learning-based clinical tools before their implementation in patient care. In line with this recommendation, we aim to develop a clinical deep learning-based, markerless lung tumor tracking framework that integrates a patient-specific quality assurance (QA) method.
Methods: During treatment planning, a patient-specific deep learning model was trained on simulation CT to enhance tumor contrast on kV projection images. An in-house clinical software deployed the model to track tumor motion on the enhanced kV projections using template matching. QA was performed through the following steps: (1) tracking the tumor on pre-treatment CBCT projections, (2) triangulating the average 3D tumor position from the 2D tracking results, and (3) validating the 3D tumor position against manual alignments from the same CBCT scan. These steps can be completed in real-time with minimal clinical workflow interruption. The method supports free breathing (FB), deep inspiration breath hold (DIBH) and gated treatments. The model passed QA if the discrepancy between the tracked and manually determined tumor positions was under 2 mm.
Results: The software was clinically applied to 9 FB and 14 DIBH lung SBRT patients across 50 treatment sessions, supervised by certified physicists. The average discrepancies between tracked and manually determined tumor positions were 0.99 ± 0.76 mm, 0.92 ± 0.83 mm, and 0.90 ± 0.65 mm in the vertical, longitudinal, and lateral directions, respectively. A 90% QA pass rate was achieved.
Conclusion: This framework provides a robust and efficient QA mechanism to verify model accuracy prior to clinical use. It ensures reliable intrafractional tumor tracking in lung SBRT, supporting confidence in radiotherapy treatments with minimal workflow disruption.