Automatic segmentation of prostate in MRI based on deep learning

Abstract

Prostate cancer is one of the most causes of cancer death for men worldwide and is still a significant public health problem. However, the prostate cancer can be cured if it is detected at early stage. Due to its ability to produce detailed anatomical structure,Magnetic Resonance Imaging (MRI) is one of the most used modality for prostate cancer diagnosis and treatment. Accurate segmentation of the prostate from MRI is crucial for diagnosis and treatment planning of prostate cancer. Deep learning has provided an important support in early disease detection, image processing and analysis, especially in image classification, image registration, image segmentation and in medical treatment plan. In this thesis, we propose an automatic segmentation of prostate in MRI based on deep learning methods. A Convolution Neural Network special type named 3D U-Net is used to segment the prostate in MRI.We conduct the 10 foldcross-validation experiments on the public Promise 12 Data set of 50 prostate images, and achieved a mean Dicesimilarity coefficient of 84.92% and a mean Hausdorff distance of 5.3 mm. The experiments proved that the proposed algorithm performed with promising results.