A Tversky Loss-Based Convolutional Neural Network for Liver Vessels Segmentation

摘要

The volumetric estimation of organs is a crucial issue both for the diagnosis or assessment of pathologies and for surgical planning. Three-dimensional imaging techniques, e.g. Computed Tomography (CT), are widely used for this task, allowing to perform 3D analysis based on the segmentation of each bi-dimensional slice. In this paper, we considered a fully automatic setup based on Convolutional Neural Networks (CNNs) for the semantic segmentation of human liver parenchyma and vessels in CT scans. Vessels segmentation is also crucial for surgical planning as it allows separating the liver into anatomical segments, each with its own vascularization. The CNN model proposed for liver segmentation has been trained by minimizing the Dice loss function, whereas a Tversky loss-based function has been exploited in designing the CNN model for liver vessels segmentation, aiming at penalizing the false negatives more than the false positives. In this work, the training set from the Liver Tumor Segmentation (LiTS) Challenge, composed of 131 CT scans, was considered for training and tuning the architectural hyper-parameters of the liver parenchyma segmentation model; 20 CT scans of the SLIVER07 dataset, instead, were used as the test set for a final estimation of the proposed method. Moreover, 20 CT scans from the 3D-IRCADb were considered as a training set for the liver vessels segmentation model and four CT scans from Polyclinic of Bari were used as an independent test set. Obtained results are promising, being the determined Dice Coefficient higher than 96% for the liver parenchyma model on the considered test set, and Accuracy higher than 99% for the suggested liver vessels model.