Automatic left atrium segmentation and volume estimate via 3D U-Net: a tool to ease RFA decision-making process

Atrial fibrillation is the most common type of arrhythmia, and it is believed to spread even more in the years to come. Radiofrequency ablation has recently proven its validity at addressing such condition and, alongside, left atrium volume has proven to be a crucial parameter in determining the outcome of this procedure. During the last years, many ways to automatically extract useful parameters from medical images have been developed, including left atrium volume and structure from magnetic resonance volumes. This project aims to develop a three-dimensional neural network, based on U-Net architecture, able to automatically segment left atrium and estimate its volume, starting from a whole magnetic resonance image with late gadolinium enhancement, simultaneously evaluating how this substantial simplification of more popular three-dimensional approaches would affect results. Binary masks identifying the left atrium were present in the dataset and served as ground truth. All things considered, it seems that the trade-offs in resolution necessary to ease the process and reduce computational burden may not be worth. Simplifying the input data at this extent, the network is not able anymore to recognize the fine structure at the edges of left atrium, and the resulting masks carry little to no clinical relevance. Further experiments should be carried, limiting the detrimental preprocessing on input data, to explore if a proper balance that allows the use of whole three-dimensional images in 3D networks exists.