Exploration of AI-Assisted On-Board Perfusion Imaging Technique on Patients Undergoing Thoraxradiotherapy

Abstract

Pulmonary function of lung cancer patients can significantly change along the radiation therapy (RT) course. Scholars have synthesized lung function maps from computed tomography (CT) images. Still, there is lack of study exploring the feasibility of generating lung perfusion information from cone-beam CT (CBCT). Our study aims to fill up this gap in the body of literatures. One-hundred-and-six pairs of planning CT and CBCT images of thoracic cancer patients from XX hospital were collected retrospectively. All CT images were registered to CBCT with a deformable algorithm, and referred as deformed planning CT (dCT). A 2D convolutional neural network (CNN) model built with cycle generative adversarial network (cycle-GAN) was trained to correct CBCT image artifacts and HU value discrepancies. The corrected CBCT was named enhanced CBCT (eCBCT) in this study. All CBCT, eCBCT and dCT images were then input to a pre-trained CT-to-perfusion deep learning (DL) model to synthesize perfusion images (PI), namely PICBCT, PIeCBCT and PIdCT respectively. For model training, 80 image pairs were assigned to the training set while others to the testing set. dCT and PIdCT were respectively regarded as the ground truth in two consecutive models. One extra validation case acquired with SPECT perfusion scan was collected for model performance test. Quantitative evaluation was done between the ground truth and the synthesized images. Peak signal-to-noise ratio (PSNR) and mean absolute error (MAE) were computed to assess efficacy of CBCT enhancement. Correlation (R) and Dice similarity coefficient (DSC) were used to evaluate voxel-wise and function-wise concurrence in CBCT-derived perfusion mapping. In CBCT enhancement, [eCBCT-dCT] pairs demonstrated higher agreement than [CBCT-dCT] pairs. Quantitatively, PSNR and MAE were improved from 21.10±1.60 to 24.08±1.76, and 68.99±13.51 to 47.06±11.31 (p<0.01), respectively. For perfusion translation, higher correlation was demonstrated in both voxel-wise and function-wise evaluation within [PIeCBCT-PIdCT] than [PICBCT-PIdCT]. R correlation increased from 0.84±0.09 to 0.89±0.06 (p<0.01), whereas DSC in high-functional regions increased from 0.77±0.06 to 0.82±0.05 (p<0.01). In the validation case, strong correlation was observed between SPECT perfusion scan and PIeCBCT, with R increasing from 0.57 to 0.65 when compared to PICBCT. We proposed a novel DL framework that synthesized perfusion images from fractionally acquired CBCT. Our framework met the clinical needs in providing real-time lung perfusion maps. It can be used to track the continuously changing pulmonary function status throughout the RT course using routinely scanned CBCT, without additional imaging modalities. The framework can facilitate functionally guided adaptive radiation therapy by providing immediate pulmonary function information.