The Development Of A Machine-learning Based System For Prostate Segmentation Of CT Images

Prostate cancer is one of the most common cancer in male reproductive system,which becomes a threat to human health.Therefore,the effective treatment of prostate cancer is clinically meaningful and important to the decline of cancer mortality in China.The image-guided radiation therapy provides effective tools for cancer therapy,which needs to accurately and automatically segment the prostates in CT(Computed Tomography)images in advance.However,the low contrast between prostate and its nearby organs,unpredicted daily prostate motions and dramatic appearance change of CT images across different treatment days make the segmentation of prostates quite challenging.Firstly,CT Prostate Segmentation Based on Continuously Updated Random ForestsTo solve the problems described above,this dissertation proposes a new segmentation method based on online update scheme and random forest.It treats the Haar-like features and labels of the sample points as input and uses auto context model to train a series of random forests for classification.These classifiers will be subsequently used to segment the prostate in the newly acquired CT images.After the new segmentation result is obtained and reviewed by clinicians,online update scheme adds the results to pool of training images to retrain the classifiers.As more and more new segmentation results are acquired during the treatment,the shape information of prostates for the current patient can be continuously incorporated into the training procedure by online update scheme.In this way,the classifiers can be updated continuously,and consequently the accuracy of prostate segmentation will be improved.The experimental results show that online update scheme can improve the performance of prostate segmentation in CT treatment images.Secondly,Multi-task Prostate Segmentation for Planning CT Images Based on Population ImagesAdditionally,this dissertation develops a multi-task CT prostate segmentation method for planning images which makes full use of population images.At first,the population images from different patients are mapped to various spaces of reference images to form a multiple training task.Then,the random forest method and auto context model are used to train a series of classifiers for each training task.After that,the classifiers acquired in each training task are applied to the CT planning image to be segmented and the final results can be obtained by using majority voting method.It can be concluded from the experimental results that since multiple training tasks can discover more information hidden in the training dataset,the proposed method can achieve higher accuracy of prostate segmentation,with the comparison to the traditional prostate segmentation method.