Treenet: Multi-Loss Deep Learning Network To Predict Branch Direction For Extracting 3d Anatomical Trees

Calculation of blood vessel or airway direction is important for the task of tree tracking in 3D medical images. However, most existing works treat branch direction estimation as only a by-product of vesselness or tubularness computation. In this work, we propose a deep learning framework for predicting tracking directions of anatomical tree structures. We modify the deep V-Net architecture with extra layers and leverage a novel multi-loss function that encodes direction as well as cross sectional plane information. We evaluate our method on both 3D synthetic and 3D clinical pulmonary CT datasets. On the synthetic dataset, we outperform state of the art methods by at least 10% in direction estimation accuracy. For the clinical dataset, we outperform competing methods by 1-4% in mean direction accuracy and 4-10% in corresponding standard deviation.