Thesis Of Shenzhen University: Overlapping Objects Splitting In Microscope Images

Comparing to other imaging methods, such as MRI or Ultrasound, microscope images can make cell-level information accessible and sufficient; thus capturing signals of disease in its early stage more detailed and accurate. In clinic, therefore, approaches based on microscope images usually are the golden standards of many diseases. However, as the big data of patent images, analyzing these image data performed by microscopist is a laborintensive task. And locating a few abnormal cells that directly contribute to identify disease is also difficult when the microscopist lacks suifficient experience. Meanwhile, the miscellaneous types of cell and the evolution of cell cycle make various appearance and heterogeneous shapes, so doctors' manual-labeled methods do not able to apply the diagnosis requirements of big data.Hence, this thesis focus on the most crucial and difficult problem of automatic analysis microscope images, including segmenting object regions, detecting overlapping regions, splitting overlapping regions and inferring overlapped parts. Our contributions are highlighted as follows:1. We propose a method based on multiple scale convolutional neural network for segmenting object regions. Differing the traditional learning-based network, our network learning feature from train sample with different scales and then learning the output of each scale's network to sgment object regions. This network can capture more context information and control model parameters effectively. The results demonstrate this method not only segments object region productively but also tunes model's parameters easily.2. We develope an approach based on image content saliency and distance relationship for segmenting object region. This approach capture the difference of different regions to locate local saliency. Next, incorporating content effects of regions with different distance. Results of this approach illustrate it is effective for segmenting objects, which have the wide of their size, in microscope image, and the computational complexity is also efficient comparing other traditional methods.3. We explore new approaches of overlapping objects splitting and thus proposed a method based on shape and color difference to detect overlapping regions. This method locate markers of corresponding overlapped objects to compute the object number in the overlapping region. Meanwhile, establishing the evidence between pixels and these markers split overlapping objects. The experiment shows this approach can compute the number of microscope images and split overlapping objects with different shape accurately.4. We propose a method to infer the shape of overlapped parts. Firstly, pixels are classified to their corresponding markers and then correcting clumps shape according to the learned shape template. Finally, involving the boundaries of split regions resort to the color information. We also demonstrate the proposed method have strong ability to recover shape of overlapped objects, and can split overlapping objects that overlapping deeply.In summary, this thesis concentrates on segmentation of object region and splitting of overlapping objects. Explored some significant issues makes microscope images analysis automatically. We also demonstrate the performance of proposed methods with sufficient experiments. Theoretically, proposed method in this paper can be applied the most of applications in automatic slide analysis. And ideas in this paper are also useful to quantitative analyse of objects' number, staining pattern and shape etc. Furthermore, the proposed theory, model and method are also beneficial for other applications of overlapping objects splitting.