SARSA-based Biomedical Image Segmentation And Its Application

Biomedical image contains a wealth of important information. But with the increasing number of biomedical images, it is impossible for human to process manually. A more feasible method is to use a computer to analyze and process images, and then to identify the structure of the organization and other interested organizations, because of the complexity and diversity of medical images, the commonly used segmentation algorithm in the application of biomedical images, the effect is poor.In this dissertation, we first analyze the biomedical images, and combine the information with the biomedical domain knowledge. We confirm that binarizing of biomedical images will not affect the actual detection results. Taking into consideration of the variation within a biomedical image and differences among biomedical images, we, taking advantage of Sarsa(State-action-return-state-action, state, action), propose a Sarsa-based image binarization approach, which separates the image into several sections according to the content of the image, and set a binarization threshold for each section a threshold. At the same time, we put forward an improved Gauss filtering denoising method to remove the noise in the image, which balances the size of the template and the size of the sliding scale parameter so as to improve the one-dimensional Gauss function. The experiment results show that our proposed method is better than the conventional threshold methods.Because most of neural EM images are of the low resolution, of low resolution and full of noise, there will be some gaps after boundary detection, which adds difficulty to the recognition of the structure of neurons. Connecting and fixing the gaps will improve the understanding of the image. While simply using of the shortest line as a connecting line is not a good choice and is possible to change the shape of the object, resulting in incorrect of the structure identification. In this dissertation, the shape of curve and the trend of image boundary are used to bridge the gap, and minimizing the difference between the fixing curve and the actual curve is the goal of optimizing. Using the original image data, we propose a new algorithm based on Sarsa, which can reduce the error rate, Rand error rate and Warping error rate.Finally, we use the Symposium on Biomedical Imaging(International) 2012(Drosophila) 30 sets of EM images to test our approach. The testing results on 30 EM images from ISBI 2012 indicate that both of our approaches, whichever the one with boundary amending or the one without boundary amending, surpassed the other boundary detection approaches, and that boundary amending really yielded an improvement eventual detection ability.