Dose reduction potential in dual-energy subtraction chest radiography based on the relationship between spatial-resolution property and segmentation accuracy of the tumor area

We investigated the relationship between the spatial-resolution property of soft tissue images and the lesion detection ability using Unet. We aimed to explore the possibility of dose reduction during energy subtraction chest radiography. The correlation between the spatial-resolution property of each dose image and the segmentation accuracy of the tumor area in the four regions where the tumor was placed was evaluated using linear regression analysis. The spatial-resolution property was determined by task-based evaluation, and the task-based modulation transfer function (TTF) was computed as its index. TTFs of the reference dose image and the 75 % dose image showed almost the same frequency characteristics regardless of the location of the tumor, and the Dice coefficient also high. When the tumor was located in the right supraclavicular region and under 50 % dose, the frequency characteristics were significantly reduced, and the Dice coefficient was also low. Our results showed a close relationship between the spatial-resolution property and the segmentation accuracy of tumor area using deep learning in dual-energy subtraction chest radiography. In conclusion, a dose reduction of approximately 25 % compared to the conventional method can be achieved. The limitations are the shape of the simulated mass and the use of chest phantom.