This study discussed the development of the FORSSMANN algorithm, a deep learning (DL) model designed to identify anatomical risk-factors on transcatheter aortic valve replacement (TAVR) candidates. FORSSMANN was trained using CT scans of 800 candidates from Fuwai Hospital, as well as internally and externally validated on a total of 525 candidates from varying hospitals. Five major anatomical risk factors correlated with adverse events during TAVR were identified, and FORSSMANN was trained to systematically detect key anatomic structures and the five risk factors using a coarse-to-fine DL network. To validate the algorithms, two experienced pre-TAVR imaging observers assessed and digitally annotated the CT images used for internal and external validation. Both validations showed a strong consistency with the senior observers’ annotations; mean Euclidean distances between manual annotations and FORSSMANN segmenting of different key structures ranged from 0.928 mm (SD 0.432) to 0.959 mm (SD 0.441) in the internal validation and 0.947 mm (SD 0.489) to 1.098 mm (SD 0.487) in the external validation. Additionally, FORSSMANN displayed high accuracy in identifying the five anatomical risk-factors, with combined internal and external accuracy values for different risk-factors ranging from 0.970 to 0.989 when compared to manual annotation. Finally, it was shown that time spent imaging was significantly reduced using FORSSMANN compared to manual annotation, with manual observers taking an average 19.50 minutes (SD 7.55) and FORSSMANN taking an average 0.86 (SD 0.21) minutes per case.