Purpose: X-ray image-guided endovascular procedures are increasingly preferred treatment modalities for cerebrovascular conditions such as strokes and aneurysms. Dose reduction without significant loss in image quality during such procedures is crucial for increased patient safety. In this work we present a dose reduction technique that uses a convolutional neural network (CNN) to restore image quality in the dose reduced images.
Methods: A 0.7 mm thick copper with a circular hole of 10 mm was used as the Region of Interest (ROI) x-ray attenuating material to differentially reduce the x-ray dose in the periphery regions, while maintaining regular dose within the ROI. The resulting dose-reduced image was provided as input to a CNN capable of deriving a mask of the ROI attenuator without any prior knowledge of its position or size within the image. To restore image quality in the dose-reduced regions first, the brightness in the dose-reduced image was equalized by logarithmically subtracting the CNN derived mask from the dose-reduced image. A spatially different recursive filtering technique was used to reduce the noise in the brightness equalized image. A higher filter weight was used in the dose-reduced periphery region thereby reducing noise but slightly increasing lag, whereas a lower filter weight was used within the ROI to maintain temporal resolution.
Results: Using the CNN derived mask with recursive filtering, image quality in the periphery regions of the dose-reduced images was successfully restored without any significant loss in visibility. A total integral dose reduction of 62% per frame was achieved.
Conclusion: The CNN was trained to extract ROI attenuator images, without prior knowledge of the ROI position or size, thus providing a robust implementation platform for the ROI dose-reduction technique during clinical interventions.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by an equipment grant from Nvidia, USA and Canon Medical Systems Corporation.
Not Applicable / None Entered.