Researchers from the Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Clinical Center, and National Center for Biotechnology Information have introduced a new method for creating synthetic X-ray images using data from computed tomography (CT) scans. The method, called Digitally Reconstructed Radiography (DRR), uses ray tracing techniques to simulate the path of X-rays through CT volumes. Unlike traditional radiographs, DRRs offer controlled and reproducible imaging conditions which are crucial for radiation therapy planning, surgical preparation, and algorithm development.
The team has applied this method to the CT-RATE dataset to create the DRR-RATE dataset. It comprises 50,188 chest CT volumes from 21,304 patients, each with a text radiology report and binary labels for 18 types of pathology. This dataset is designed to support research in disease detection and AI applications in the fields of radiology and other related disciplines.
The DRR-RATE dataset is derived from CT scans, and it offers labeled images and radiological reports. By simulating CT-derived pathologies in X-ray form, DRR-RATE enriches training data for diagnostic models and enhances our understanding across imaging modalities. This dataset is publicly accessible under a CC BY-NC-SA license.
In experiments with the DRR-RATE dataset, the researchers trained and evaluated the CheXnet model for chest X-ray classification. The model demonstrated notable results for the classifications of Cardiomegaly and Pleural Effusion with area under the ROC curve (AUC) scores of 0.92 and 0.95, respectively. However, classifications of Atelectasis and Consolidation were less robust with AUC values of 0.72 and 0.74. Classifications of Lung Nodule and Lung Opacity had even lower AUC scores outlining room for improvement.
When CheXnet was trained on a different dataset, CheXpert, and tested on DRR-RATE, performance decreased slightly for most conditions, which was attributed to the domain differences between real and DRR images. Hence, while the DRR-RATE technique shows promise, it also highlights the challenges ahead for less prominent conditions like Atelectasis, Lung Nodule, and Lung Opacity, potentially due to resolution limitations in DRR images.
Overall, the integration of DRR-RATE signifies a considerable advancement in the synthesis of medical imaging data, bolstering AI-driven diagnostic capabilities and offering substantial potential to enrich medical research. This work also demonstrates the immense potential and capabilities of utilizing AI in the medical imaging field, and the necessity of large, diverse datasets for accurate and robust machine learning model training.
