Medical Images Sequence Normalization and Augmentation: Improve Liver Tumor Segmentation from Small Dataset

Using deep learning for Medical Images Diagnosis automatically is a new trend in recent years. Concretely, the fundamental step in these studies is the automatic segmentation system for human organ, that has the benefit of accuracy in the diagnosis process for medical images (CT images). Traditionally, if there is no module of automatic segmentation system, this process has to be performed by the experience of specialized physicians. More importantly, this process takes much time of physicians, while the automatic segmentation system can alternative efficiently. However, the accuracy and responsiveness are the challenges for a segmentation system with a small organ or small tissue, i.e., a liver tumor, because of the lack of data for deep learning. This paper highlights some augmentation and preprocessing techniques from a small dataset of medical images. We propose a method to improve the module of automatic segmentation, in particular, liver tumor segmentation. For the augmented dataset from 3D Image Reconstruction for Comparison of Algorithm Database (3D-IRCADb), we achieve the dice score of 94.58% for liver segmentation and 45.87% for liver tumor. Furthermore, with a real dataset from a patient, the dice score of our system is over 78% for liver segmentation.