Local integration of deep learning for advanced visualization in congenital heart disease surgical planning

Artificial intelligence (AI) based convolutional neural networks (CNN) can help clinicians to make faster blood pool segmentations, which are useful for new technologies such as 3D printing, augmented reality applications, etc., to assist surgical planning. We report first experience from integration and adoption of deep learning to local needs in the field of congenital heart surgery. We propose a faster way of creating clinically approved segmentations, based on publicly available computed tomography (CT) images. The DenseVNet CNN with multiple loss functions is trained on a publicly available dataset with 66 CT volumes, which has been used to create blood pool segmentations for our local hospital dataset. AI-assisted blood pool segmentation led to a significant reduction in segmentation time of 36 ± 12 min (mean ± standard deviation), compared to semi-automatic methods of 203 ± 133 min. We have obtained the Dice score = 0.9183 ± 0.0351, oversegmentation = 0.1050 ± 0.0972 and undersegmentation = 0.0584 ± 0.0462 for the public test dataset. The local hospital test dataset shows the Dice score = 0.8595 ± 0.0733, oversegmentation = 0.0518 ± 0.0315 and undersegmentation = 0.2343 ± 0.1574. We observed a significant improvement in the Dice score compared to state-of-the-art methods for the public dataset. The inclusion of over-and undersegmentation in the evaluation metrics gives a better clinical understanding along with the Dice score. Further, we have printed a 3D model of blood pool segmentation for educational purposes for clinicians under training. The workflow of combining deep learning and manual annotation is useful to create more blood pool segmentations which we use for research on 3D visualization methods in surgical planning.