Using artificial intelligence for device identification and characterization in angiographic sequences of TAVI procedures as radiomic biomarkers for prosthesis deterioration

INTRODUCTION: Transcatheter Aortic Valve Implantation (TAVI) has been established as the preferential option for patients suffering from symptomatic severe aortic stenosis with high surgical risk. Due to the implantation of prosthesis on younger patients, the durability of TAVI devices becomes a crucial issue, especially beyond 5 years of follow-up [1]. Therefore, an application of growing interest is the assessment of the prosthesis deterioration, which is directly related to the measurement of the stent tip deflection [2]. In order to extract this parameter from angiographic sequences, a prior requirement relies on the automatic detection of the prosthesis in the image, pointing directly to the use of Artificial Intelligence. PURPOSE: Automatically segment TAVI prostheses in angiographic sequences using Deep Learning [3]. The aim is allowing the extraction of information from these images, as the measurement of the stent tip deflection for evaluating the prosthesis deterioration, to improve the clinical practice. METHODS: A U-Net, which is a Deep Learning architecture designed for the analysis of biomedical imaging, has been trained for the segmentation of TAVI prostheses in angiographies [3]. The self-built dataset includes 50 sequences captured during TAVI from a population of 15 patients, obtaining 2827 frames (examples provided in Fig. 1). This dataset is randomly split into training and test sets using an 80/20 ratio. The U-Net is trained using the training set and its annotations (manually generated), optimised during 50 epochs using Adam optimiser, and binary cross-entropy as the cost function. Once the model has been trained, it is used for the segmentation of the images in the test set. RESULTS: The segmentation results have been evaluated in separate subsets of frames extracted from the test set, regarding the stage of the procedure they correspond. Such subsets are aortic root previous to the device delivery (Fig. 1A), prosthesis deployment (Fig. 1B), prosthesis fully-expanded (Fig. 1C), and contrast agent injection (Fig. 1D). Fig. 2 provides examples of the results obtained for the images in Fig. 1, showing the original frames and the segmentation result overlaid. The results have been evaluated in terms of different classification metrics, obtaining the mean values in the following ranges for the four groups: accuracy 0.99–1.00, recall 0.70–0.97, and area under the receiver operating characteristic curve (AUROC) 0.85–0.98. CONCLUSION: U-Net has been successfully applied to angiographic imaging captured during TAVI procedure for segmenting the prosthesis, with AUROC values between 0.85 and 0.98. The prosthesis identification and characterization allow the extraction of imaging biomarkers of great interest that can be used for the evaluation of the prosthesis deterioration. FUNDING ACKNOWLEDGEMENT: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Axencia Galega de Innovaciόn (GAIN) (Code Numbers IN845D-2020/29 and IN607B-2021/18)