Computer-Aided Detection For Brain Tumor

Accurate tumor segmentation is an essential and crucial step for computer-aided brain tumor diagnosis. It can provied doctors with a basis for surgical planning. It also can ensure the brain tumor's resection more thoroughly while protect the normal tissues. Moreover, it can improve patient quality of life and prolong survival period. Manual segmentations are widely adopted in clinical diagnosis and treating, but this tradiontal way is neither accurate nor reliable. An automatical and accurate system for brain tumor segmentation is strongly expected. Automatic tumor segmentation for medical images has been studied by previous works. But they are still facing on some difficulties, such as unsatisfactory segmentation accuracy or requiring the manual intervention.In this thesis, we first introduce some medical image processing technologies, and then introduce the concept of deep learning. Based on the realted work, this thesis proposed a corse-to- fine method to segment the brain tumor. This proposed hierarchical framework consists of preprocessing, deep learning network-based segmentation and post-processing. The preprocessing part mainly includes the medical image denoising, brightness transformation and histogram processing so that the brain MRI images under different instrument parameters can be more consistent contrast and brightness. Then, we extract image patches for each MRI image, record the original coordinates of the center point for each image patch, and obtain the gray level sequences of image patches as the input of the deep learning network so as to avoid to extract additional features.The deep learning network based segmentation is implemented by a stacked denoising auto-encoder network. At firstly, it utilizies a greedy layer-wise initialization procedure to obtain the original parameters of the deep ne twork, then attempts to fine-tune all parameters of this deep architecture by employing the backpropagation. The hierarchical structure can help users to extract the high level abstract feature from the input, and utilize these absatct features to classify image patches.After mapping the classification result to a binary image, a series of morphological filters are implemented to get the final segmentation result. We call these steps as post-processing.Finally, the proposed method was applied to segment the brain tumor on the real patient MRI dataset. By evaluating the experiment, the final segementation results show that the proposed method is more accurate and efficient.