| Summary: | 碩士 === 國立聯合大學 === 電機工程學系碩士班 === 106 === Clinically, the judgment of the tumor needs to be interpreted by the pathologist for the pathological tissue section. In addition to providing information on tumor size and type, pathological tissue section also can assist physician to choose the appropriate treatment. For this reason, correct interpretation of pathological sections is crucial. However, pathologists mostly analyze pathological sections manually. Long-term observation of microscopic images caused waste of manpower, waste of time, and misjudgment by pathologists. Therefore, developing a computer-aided diagnosis system for correctly detecting all cells on the pathological section is one of the current clinically important needs.
Traditionally, in order to count cells we need to use image segmentation algorithms to find the cell contour first. In the process, there will be problems such as overlapping cells, uneven staining, and inconspicuous cell boundaries. Deep learning can more effectively detect cells through artificial intelligence. This study attempts to use the deep learning model, Faster-RCNN, to apply to different stained pathological sections for cell counting. The feature matrix extraction network of Faster-RCNN uses the convolution layer of AlexNet and VGG16 respectively. We adjusted the parameters and used the original image of the pathological tissue section for network training.
Training two networks, AlexNet as a convolutional layer of Faster-RCNN, and VGG16 as a convolutional layer of Faster-RCNN. The results show that Ki-67 stained slice cell detection with the existing deep learning network can achieve 90% accuracy, and more importantly, the computation time can be greatly reduced. This method drastically reduces the computation time, including training phase and testing phase. This study confirms the feasibility of using the same deep learning network for cell counting in Ki-67 pathological sections. We expect to build a more efficient cell counting method by adding image pre-processing and modifying the network architecture.
|
|---|
