Computed Tomography, Pancreas segmentation, Convolutional Neural Network, Long Short-Term Memory
Abstract
Compute Tomography (CT) imaging is mostly used to diagnose abdominal disease. Among many organs in abdominal CT, the pancreas is one of important regions to diagnose diabetes, pancreatic cancer, and pancreatitis. Thus, it is important to fin the pancreas area and quantitatively analyze the present of disease. However, it is difficult to segment the pancreas area due to the size and shape variation depending on patients and ambiguous boundaries with surrounding organs. A lot of methods have been proposed for the automatic pancreas segmentation, but the accurate segmentation is still challenging. Recently, deep learning methods have been achieved better performance than conventional machine learning based segmentation methods for many applications. Thus, in this paper, we propose a new pancreas segmentation method using deep learning. Applying a deep learning algorithm to the 3D CT image segmentation is non-trivial due to a memory limitation. Several deep learning methods address this issue by dividing a 3D medical image into small 3D patches and then applying 3D Convolutional Neural Networks (CNN) on the patches. However, anatomical information cannot be considered in this way. To address these issues, in this paper, we propose a Multi-Dimensional Long Short-Term Memory (MDLSTM) method which can consider the anatomical information using a sequential learning scheme. MDLSTM is compose of a 3D CNN that performs the patch-wise segmentation on local regions and a Long Short-Term Memory (LSTM) that propagates information of adjacent patches. For evaluation, the NIH CT dataset was used to train the MDLSTM model, and Dice-Sorensen Coefficient (DSC), Precision, and Recall were used to compare performance with other 2D and 3D CNN methods. When using the proposed method, we can see the DSC and precision were increased by 2%, 16%, respectively compared to the conventional CNN method.
Table Of Contents
List of Contents Abstract ·································································································· i List of contents ························································································· ii List of tables ··························································································· iii List of figures ·························································································· iv Ⅰ. Introduction 1.1 Introduction ·················································································· 1 1.2 Related works ················································································ 2 Ⅱ. Method 2.1 U-net ·························································································· 4 2.2 Convolutional Long Short-Term Memory ··············································· 5 2.3 Multi-Dimensional Convolutional Long Short-Term Memory ························ 6 2.4 Forward and Backward propagation for the MDCLSTM network ···················· 7 Ⅲ. Experimental results 3.1 Dataset ························································································ 8 3.2 Implementation details ······································································ 8 3.3 Evaluation metrics ··········································································· 8 3.3 Qualitative and Quantitative Analysis ···················································· 9 Ⅵ. Conclusion ······················································································· 12 Reference ······························································································ 13