269 - Markerless Dynamic Tumor Tracking Using YOLO-Based Detection on Rotated Orthogonal Fluoroscopic Images Training with 4D-CT-Derived DRRs

Purpose/Objective(s):

For dynamic tumor tracking, implanted markers are commonly used to facilitate tumor localization. However, marker implantation is invasive and may lead to complications such as marker migration or loss. To address these limitations, we developed an AI-based framework for direct tumor detection on rotated fluoroscopic images trained using rotated digitally reconstructed radiographs (DRRs) generated from four-dimensional computed tomography (4D-CT).

Materials/Methods:

Thirteen lung cancer patients with respiratory motion who underwent simultaneous acquisition of rotated fluoroscopic images from orthogonal directions were included. The median tumor motion in the superior-inferior (SI) direction ranged from 8.7–37.9 mm with a median of 16.0 mm. Gross tumor volume (GTV) ranged from 0.5–41.77 cm³ with a median of 7.87 cm³.

Rotated DRRs were generated from 4D-CT for six respiratory phases (0%, 20%, 40%, 50%, 70%, and 90%) within the same angular range as fluoroscopy at 2-degree intervals. To enhance contrast and augment training data, gamma transformations using eight gamma values (1/10, 1/8, 1/6, 1/4, 1/2, 1, 2, and 3) were applied.

GTV-only DRRs were additionally generated from GTV contours, and the projected GTV masks were extracted. Pairs of gamma-transformed DRRs and the corresponding GTV contours were used to train patient-specific YOLO (version 11) models.

For testing, tumors were detected on fluoroscopic images from each view using the Slicing Aided Hyper Inference method, and three-dimensional (3D) positions were calculated by triangulation from orthogonal directions. Detection rate was defined as the proportion of image pairs with successful bilateral detection and 3D localization.

Reference tumor positions were estimated by adding respiratory motion-based offset vectors to the centroid of implanted markers, with respiratory phases determined from infrared signals. These pseudo tumor positions served as ground truth. Localization errors were evaluated in the LR, SI, AP directions and 3D distance, and median 3D error was calculated for each patient.

Results:

The median detection rate across all patients (IQR) was 97.4% (78.2–100.0%). The median localization errors pooled across all patients (IQR) were 0.85 mm (0.41–1.58 mm), 0.86 mm (0.42–1.57 mm), 0.87 mm (0.38–1.63 mm) in LR, SI and AP directions, respectively, with a median 3D error of 1.98 mm (1.33–2.83 mm). Patient-specific median 3D localization errors ranged from 1.34–4.86 mm, with an overall median of 2.23 mm.

Conclusion:

In this study, a direct tumor detection algorithm for rotated fluoroscopic images was developed by training YOLO using gamma-transformed rotated DRRs as contrast-augmented data. The proposed approach achieved high detection rates with sub-2-mm median 3D localization accuracy, demonstrating the feasibility of markerless dynamic tumor tracking. These results suggest the potential applicability of the proposed framework to motion-managed VMAT.