Deep learning model for 3D computed tomography (CT) image reconstruction with single or a few views

Stanford researchers have developed a novel and efficient method for generating real-time 3D volumetric computed tomography (CT) images with 2D single or few-view projections, instead of several hundreds of projections as required in existing CT imaging system. This invention completely alleviates the need of angular sampling required by conventional CT imaging, saving time and costs. The deep learning model uses pre-treatment patient data to train the model to transform 2D projections into 3D images and construct a robust encoding/decoding framework.
With online, real-time single projection acquisition, applications include single-view 3D image-guided radiation therapy as well as other image guided interventional procedures such as high intensity focused ultrasound, surgery, biopsy, cardiovascular procedures, and radiofrequency treatments.

Figure

Figure description - Architecture of the proposed deep learning network for single-view and few-view CT image reconstruction

Stage of Research