Relationship between task-based modulation transfer function and evaluation index of tumor area in dual energy subtraction chest radiography

Abstract

When extracting lesions in medical images using deep learning, the extraction accuracy will be higher if the image quality is good. The purpose of this study is to investigate and comprehend the relationship between the spatial-resolution property of X-ray images and the accuracy of lesion extraction by deep learning, and to examine the possibility of dose reduction from the obtained results. The simulated masses arranged on different areas in dual energy subtraction (DES) chest radiographs acquired by various doses were used for the investigation. Task-based transfer functions (TTFs) as a spatial-resolution property for the DES chest radiography were calculated. The mass areas were also extracted by U-net, and Dice coefficients were obtained as the extraction accuracy. In results, regardless of mass locations, the TTFs of the reference dose images and the 75% dose images showed high frequency responses, and the Dice coefficients were also high. The TTFs and the Dice coefficients were obviously lower in the images of when the masses were located in the right supraclavicular region at 50% dose compared with the other conditions. The results of this study suggested that the spatial-resolution property was strongly related to the accuracy of mass region extracted by deep learning in DES chest radiography. In conclusion, the dose reduction of about 25% compared with the conventional dose should be feasible.

Keywords

• Deep learning
• Spatial-resolution property
• Dice coefficient
• 1-Shot dual energy subtraction