Cone Beam Computed Tomography-Based Lung Function Imaging Using Deep Learning for Functional Image Guided Adaptative Radiation Therapy.

PURPOSE: Recent studies indicate that lung function can change significantly during the radiation therapy (RT) course. However, additional functional imaging scans are not part of routine RT workflow, limiting the use of functional avoidance for adaptive therapy. To bridge this gap, a deep learning model was developed to synthesize functional maps directly from fractional cone beam computed tomography (CBCT) images, enabling potential functional image guided adaptive RT.

METHODS AND MATERIALS: Data were prospectively collected from 60 lung cancer patients who underwent intensity modulated RT. In addition to standard planning computed tomography and fractional CBCT scans, all patients received a baseline single-photon emission computed tomography (SPECT) perfusion scan before RT. A subset of 16 patients also underwent a follow-up SPECT scan after completing RT. A 3-dimensional Gated-Attention U-Net was developed to synthesize perfusion maps directly from CBCT images. To address the challenges of inherent CBCT noise and artifacts, the network architecture was augmented by integrating gated attention modules and 3-dimensional deformable convolutions within the skip-connection pathways. This design enhances multiscale feature fusion for more robust image synthesis. The synthesized perfusion images were quantitatively compared to the reference SPECT scans using cross-validation. Voxel-wise agreement was assessed using the Spearman rank correlation coefficient (R), structural similarity index measure, and mean squared error, while functional region agreement was evaluated using the Dice similarity coefficient. The potential clinical benefit was assessed through dosimetric evaluation.

RESULTS: Quantitative analysis demonstrated strong agreement between the CBCT-based functional images and the ground-truth SPECT images, with an R of 0.68 ± 0.09, structural similarity index measure of 0.75 ± 0.09, mean absolute error of 0.15 ± 0.03, mean squared error of 0.04 ± 0.02, and Dice similarity coefficient values of 0.76 ± 0.07 and 0.85 ± 0.05 for high- and low-functional regions, respectively. Regarding dose sparing of the lung high-function region, the CBCT-based functional image guided plan significantly reduced the mean dose by 6.97 ± 4.22 Gy and V20 by 13.54% ± 8.66% compared to the anatomic plan, while achieving a better target dose homogeneity index of 5.26 ± 0.65.

CONCLUSIONS: A CBCT-based lung function imaging method was developed using a deep learning model. The evaluation demonstrated its feasibility for functional guidance planning in functional image guided adaptive RT. A larger cohort study is warranted in the future.