| With the development of economic and the improvement of people’s living standard, people begin to pay more and more attention to their own health problems. Prostate cancer is the most common cancer in the male reproductive system. In Europe and other countries, the incidence of prostate cancer in men ranked first in cancer rates. With the development of the medical imaging technology, because many of the advantages of magnetic resonance images are applied to the diagnosis and treatment of prostate cancer. How to segment the prostate effectively and efficiently become an urgent subject for computer aider medical diagnosis.Image representation plays an important role in medical image analysis. How we represent the input data is the secret of success to different medical images algorithms. In other words, how do we use features to characterize the input image. In the past ten years, features engineering remains as a hot research field of medical images. Many innovative hand-crafted features were designed such as Haar wavelet, local binary patterns. But those hand-crafted features are inability to extract and organize salient information from the data at hand. Those hand-crafted features representation power can be varied across different patient datasets. Due to the considerable individual prostate magnetic resonance images’ difference, those features can not represent those prostate images effectively. The most effective features should be designed in a learning based manner, which can be adopted to different patient datasets at hand. Therefore, based on the problem of the automatic prostate segmentation in MRIs, we have mainly completed the following research:We introduce a deep learning framework to achieve this goal. In particular, a deep belief networks is used to learn the most effective features in a hierarchical and unsupervised manner. The learned features are adapted to the dataset at hand and encode high level semantic anatomical information and we fine-tune the network using labeled data in a supervised manner, then we utilize the learned features as the input to the softmax regression classifier, thus achieving the aim of segmentation of the data.We propose a segmentation methods based on double-deep belief network for segmenting 3d prostate MRI in the prostate MRI mode. According to the distribution of gray scales in prostate MRI, we split the training data into uniformly distributed class and non-uniformly distributed class, which will be used to train two separate but with identical parameter settings respectively. Results coming from the two networks will be combined to achieve a better segmentation results. |
