Machine Learning for Liver and Tumor Segmentation in Ultrasound Based on Labeled CT and MRI Images

Using machine learning methods to segment liver and tumor in ultrasound (US) images has been challenging due to the general lack of quality labeled US training data. This paper applies the multi-modality image fusion labeling (MMIFL) method to transfer liver segmentation labels from Computed Topography (CT) and Magnetic Resonance Imaging (MRI) volume data to the corresponding US image set. The resulting labeled US data is then used to train and test specific machine learning models. The results show that segmenting liver and lesion on the US images based on the transferred labeling from CT/MRI images is feasible. Furthermore, due to the nature of the fusion technique, transfer learning is easily implemented. For baselining, we first reproduce the public available 2D UNet tuned for the CT/MRI segmentation, and further test other model architectures. The best model had an average liver DICE score of 82.09%.