Semantic Segmentation of Liver Tumor in Contrast-enhanced Hepatic CT by Using Deep Learning with Hessian-based Enhancer with Small Training Dataset Size

摘要

Deep learning requires a large dataset for training, but collecting and annotating such a large dataset are time-consuming and often very difficult. In this study, we proposed a 3D massive-training artificial neural network (MTANN) incorporated with a Hessian-based enhancer to achieve a high performance of our MTANN model with small training datasets. Contrast-enhanced CT scans of 42 patients with 194 liver tumors from the Liver Tumor Segmentation (LiTS) Benchmark database were used in this study. MTANN models were trained with: 14 and 28 tumors. The remaining 24 patients with 59 tumors were applied to the trained models for evaluation. Our method achieved a Dice of 0.703 with a training set of 14 tumors. The performance was comparable to the best performance in the MICCAI 2017 competition with less than one tenth of training cases. Our method would be essential in applications where a large training set is not available.