External beam radiation therapy (EBRT) is an important treatment modality for lung cancer with stereotactic body radiation therapy (SBRT) now considered the standard of care for inoperable early-stage non-small cell lung cancer (NSCLC) patients. A major limiting factor to conform the high dose volume to the target is the inherent presence of positional uncertainties, particularly those caused by respiration. Respiratory motion is complex and can be affected by numerous factors, both patient- and tumor-related. Conventionally, large margins are used to ensure tumor dose coverage, inevitably resulting in increased exposure of surrounding healthy tissue. Besides, the radiation field can still miss the target because of the inherent presence of interand intrafraction variations. Ideally, the tumor position should be monitored directly at treatment. A possible solution to this is through surgically inserted radio-opaque markers. However, this still does not directly track the tumor and is error-prone due to marker migration and uncoupled motion between marker and tumor. Moreover, this approach may also not be suitable for all patients due to its invasiveness. This thesis presents the development and clinical evaluation of a markerless detection and position estimation algorithm for lung tumors, intended to serve as an independent and complementary system for intrafraction monitoring in non-invasive lung cancer EBRT. This work consists of two parts: The first part presents a novel method designed based on the concept of feature dense maps and scale-invariant feature transform (SIFT) for robust frame description and matching. The algorithm operates on fluoroscopic megavoltage (MV) projection images captured with the electronic portal imaging device (EPID) during treatment delivery. The tracked target is represented by a gridded set of features, whose local appearance and geometric properties are parameterized in the form attributes and used to find their correspondence on continuous frames. The algorithm is retrospectively evaluated on phantom motion data as well as sample clinical data acquired during lung SBRT on a gimbaled linear accelerator (LINAC). The root mean square error was determined to be < 1.2 mm for phantom and < 1.8 mm for the patient data. Using dense maps ensured a thorough coverage of the dynamic scene in the beam’s-eye-view (BEV). More importantly, it gave the handles to overcome partial occlusions of the scene by using inferences from unoccluded features. The second part of the thesis validates the utility of the MV tracking algorithm in a larger clinical data set corresponding to realistic lung cancer patients undergoing markerless dynamic tumor tracking (MLDTT). The concept of dense feature tracking to overcome anatomy- and multi-leaf-collimator (MLC)-related obstruction/occlusion is investigated using tumor cases of different sizes, motion amplitudes, and locations in the lungs. An extension to the initial implementation enabled automatic segmentation of the dynamic field aperture, which formed the basis for prospective MV verification simulations during three-dimensional conformal radiation therapy (3DCRT) and intensity-modulated radiation therapy (IMRT) deliveries. It is shown that the MVtracking information can be sufficient to quantify residual errors and the geometric uncertainty of clinical MLDTT treatment. The two-dimensional (2D) MV-measured geometric accuracy of clinical MLDTT treatment was 1.2 ± 0.6 mm and 1.7 ± 0.9 mm for the left-right (LR) and superior-inferior (SI) directions, respectively, compared to 1.1 ± 0.5 mm (LR) and 1.5 ± 0.8 mm (SI) based on ground truth manual tracking. The MV tracking algorithm showed a great potential for intra-treatment verification, which may as well be transferable to other conventional platforms.
Marco Serpa (Thu,) studied this question.