Patch-Based Deep-Learning Model With Limited Training Dataset for Liver Tumor Segmentation in Contrast-Enhanced Hepatic Computed Tomography

The automatic segmentation of liver tumors plays an important role in the diagnosis and treatment of liver cancer. While deep convolutional neural network (DCNN) models are widely used for segmentation tasks in medical imaging, they require 1,000 to 10,000 annotated cases for effective training. However, assembling datasets of this size is a significant challenge due to the labor-intensive nature of tumor annotation, which often requires the expertise of radiologists. We propose a multi-scale Hessian-enhanced patch-based neural network, which we call an MHP-Net, for liver tumor segmentation with a limited dataset. Our approach involves sampling 3D patches from the input images for training a neural network, rather than using all input images, which are commonly used in DCNN training. We applied a multi-scale Hessian ellipsoid enhancer to extract multi-scale features of the liver tumor. We implemented a region-stratified sampling strategy to prevent overfitting in patch-based neural network training. We evaluated the effectiveness of our model using a dataset from the Liver Tumor Segmentation Benchmark (LiTS). To investigate the performance of the model under limited sample-size conditions, we trained it and state-of-the-art (SOTA) deep learning models with 7, 14, and 28 tumors. Our model achieved average Dice scores of 0.691, 0.709, and 0.719 which were higher than those ranging between 0.395 and 0.641 with the SOTA models. Remarkably, our model also achieved a Dice score (0.709) on par with the top model (0.702) in the MICCAI 2017 worldwide competition, despite utilizing only 1.5% (14 out of 908 tumors) of the training data.