Room: Exhibit Hall | Forum 8
Purpose: To investigate if there are tangible differences in digital subtraction angiography (DSA) images obtained using interventional fluoroscopy systems from different vendors.
Methods: IRB approval was obtained. DSA images from a set of patients with metastatic uveal melanoma who were receiving repeat chemo/immuno-embolization treatment of the same liver lobe by the same physician using two interventional fluoroscopy systems - AlluraClarity (Philips) and Artis Q (Siemens) systems - were extracted from PACS. These images corresponded to the same anatomical field of view and were obtained from the same imaging sequence at two different time points; one, when the procedure was performed using AlluraClarity system and the other, when the procedure was performed using Artis Q system. These images were converted to JPEG format. Machine learning was performed on the commercial platform Cloud AutoML Visionbeta (Google LLC, Mountain View, CA). 1832 images from the AlluraClarity system and 1902 images from the Artis Q system were included in the analysis. From this data set 3021 images were used for training, 385 images for validation (internal hyperparameter optimization), and 328 images were used for testing (161 images from AlluraClarity system and 167 images from Artis Q system).
Results: 161 test images from AlluraClarity system (100%) were identified correctly by the trained network. Out of the 167 images obtained using the Artis Q system, 164 images (98.2%) were identified correctly by the trained network.
Conclusion: There are tangible differences in DSA images of the same patient/anatomy obtained using two different interventional fluoroscopy systems that were identified using a trained neural network. Such differences may have clinical implications while monitoring treatment response or evaluating image quality subjectively.
Funding Support, Disclosures, and Conflict of Interest: Research grant from Philips Healthcare