Digitally Reconstructed Radiograph Generation for Enabling AI/ML in Medical Imaging
Anand P. Bora, Amit D. Joshi, Suraj T. Sawant · 2020
Now-a-days artificial intelligence has huge impact on medical field and outcompeted the conventional methods that are being used. In clinical fluoroscopy procedures most of the annotated data is not archived. Hence there is lack of meaningful labelled data in radiography. To generate this data from three dimensional annotated computed tomography volume using ray casting is known as Digitally Reconstructed Radiograph (DRR). This radiograph should be very similar to the real one by adding functionality such as attenuation and different parameters. This parameters are similar to the actual system pipeline of generating x-ray. The parameters include kV, mA, rotation and translation. The pipeline also consists of scatter estimation, dynamic range management, noise generation, etc. Multiple DRRs can be generated using the proposed pipeline with given transformations and learning targets. This work focuses on improving quality and speed of DRR generation. The notable amount of diverse data can be retrieved by applying data augmentation on the given small data. Thus this DRRs can be used for training, testing and validating different task based specific AI/ML models. These DRRs can be used for various tasks in x-ray imaging as classification, anatomical landmark detection, vessel segmentation, etc. The proposed solution in this work has taken 1.2 seconds on average to generate DRR of image size 1000*1000 pixel. This implies the speed of writing as 1.5 µs per pixel.