Research On Marker-Less Tumor Segmentation Algorithm For X-Ray Fluoroscopic Image Of Lung Cancer

Tumor location changes with the patients’ respiration during radiotherapy,leading to a decrease in accuracy.Sync Tra X,a real-time tumor tracking system can perform continuous X-ray imaging of the irradiated site during treatment to help determine the tumor location and achieve high precision radiotherapy.However,problems such as tumor blurring,poor visibility and bone occlusion in captured color fluoroscopic images require clinical implantation of fiducial markers.The implantation process brings associated surgical risks.In order to achieve a real-time tumor tracking without relying on such markers,this paper studyed a tumor segmentation method for marker-less lung cancer color fluoroscopic images and constructs a corresponding image processing experimental platform.The training data used in this paper were derived from 4D-CT and color fluoroscopic images captured during radiation treatment of four patients who underwent respiratory gated radiotherapy with Sync Tra X system between 2016 and 2019.After appropriate preprocessing,three types of images were generated: bone suppressed digitally reconstructed radiographs(DRR),bone enhanced DRR,and tumor labels generated using the main part of tumors.The entropy-based adaptive channel correction algorithm proposed in this paper was used to enhance and transform the color fluoroscopic images into grayscale,which were then converted to enhanced DRR style fluoroscopic images by histogram specification.To address the problems of tumor blurring,low visibility and bone occlusion in color fluoroscopic images,a U-Net-based cascade model was proposed to perform tumor segmentation.The cascade model contains three patient-specific U-Net models: U-Net1,U-Net2 and U-Net3,with function of domain transformation,bone suppression and tumor segmentation,respectively.All models use the original U-Net architecture,with U-Net1 and U-Net2 replacing the last 2 layers with a regression layer.Color fluoroscopic images with corresponding enhanced DRR style fluoroscopic images were used to train U-Net1;bone enhanced DRR with corresponding bone suppressed DRR were used to train U-Net2;bone suppressed DRR and corresponding tumor labels were used to train U-Net3.The ratio of training set to validation set was 9:1.In this paper,an image processing experimental platform was developed using MATLAB.The three models were evaluated using with validation set,and the results are presented as mean ± standard deviation.For U-Net1,peak signal-to-noise ratio(PSNR)≈ 31.96 ± 1.94,structural similarity(SSIM)≈ 94.60 ± 1.21(%),mean square errors(MSE)≈ 57.93 ± 23.32;for U-Net2,PSNR ≈ 26.56 ± 1.72,SSIM ≈ 5.14 ± 1.48(%),MSE ≈ 158.52 ± 63.53.For U-Net3,accuracy ≈ 99.37±0.27(%),the intersection over union(Io U)≈ 82.82±7.26(%),and the recall ≈ 88.98±7.54(%).the image processing experimental platform implemented functions such as correlation operations on CT data,DRR and tumor label generation and visualization.The tumor segmentation method of marker-less lung cancer fluoroscopic images established in this paper obtained better tumor segmentation results than traditional methods or single neural network models,which can provide strong support for individualized localization and treatment,with clinical application prospects and important reference value for improving the accuracy of lung tumor radiotherapy.