Pre-operative planning of minimally invasive mitral valve surgery using deep-learning

Aim: We propose a novel pre-operative planning approach based on Deep-Learning techniques in the context of minimally invasive mitral valve surgery (MIMVS), for the identification of the mitral valve and optimal thoracic working port positioning in a patient-specific fashion. Methods: We used supervised Deep-Learning for the processing of contrast-enhanced computed tomography (CT) scans. Our algorithm consisted of four steps: segmentation on CT scans, localization of the mitral valve, creation of maps under three criteria (distance/angle between the mitral valve plane and candidate working port spots, absence/presence of ribs), and selection of optimal working port. Results: We compared the performance of the Deep-Learning-based approach vs. the previously described semiautomatic method and conventional user’s planning (Dice mean value: 93.59). The Deep-Learning method outperformed the semiautomatic method [intraclass correlation coefficient (ICC) = 0.206]. We defined two interfaces to navigate among candidate working ports. Conclusion: We suggest that the Deep-Learning-based approach may help the surgeon in identifying the most appropriate working port (thoracic access) for MIMVS, and to comparatively predict the features of different candidate thoracic accesses in individual patients, and help address individual anatomic issues. It may also help standardize the pre-operative planning and obtain a faster learning curve for trainee surgeons.