Academic sharing | design and development of multi staining pathological image information evaluation system based on openvino

Academic sharing

2022 year 6 month 11 Japan , By Beijing University of Aeronautics and Astronautics 、 Tsinghua University 、 The biomedical engineering graduate academic forum jointly held by the three universities of Beijing University of Technology —— The sub Forum on biomedical imaging and information processing was successfully held . Forum , Keynote speaker Ji Junyu commented on his topic be based on OpenVINO Design and development of multi staining pathological image information evaluation system based on Made a detailed sharing .

Using a variety of stained pathological images to manually judge the type and severity of disease is a very time-consuming 、 Laborious and subjective process . Accurate evaluation of pathological information in multi staining images is crucial to quantify the weight of various staining in the diagnosis of specific diseases .

The study is based on Convolution Siam network Developed a new deep learning method , The method The information quantification task is transformed into a similarity measure between pathological patterns and non pathological patterns on histopathological images . This method includes 5 Including pathological staining images 3 It is verified on independent data queues , It can significantly improve the accuracy of multi staining pathological image diagnosis .

OpenVINO It is a high-performance solution developed by Intel for deep learning deployment at the edge or in the cloud , It can accelerate the optimization of the model . This project is based on OpenVINO Platform to build a multi staining pathological image information evaluation system , Realize in CPU Fast reasoning of terminal , Create a foundation for clinical application .

Intel at the sharing meeting OpenVINO Platform , Shared the use of OpenVINO Experience of model optimization and reasoning acceleration on platform . The time statistical comparison of reasoning in different frameworks shows ,OpenVINO stay CPU It can significantly reduce the reasoning time of the model .

Teachers and students participating in the forum actively discuss , The lecturer also answered your questions one by one ：

Q： Whether there will be imbalance between positive and negative samples ？

A： In the training stage, the imbalance of positive and negative samples can be eliminated by artificially controlling the amount of data ; And in the reasoning stage , In order to deal with the diagnostic difficulties caused by only a small number of lesions in large-scale full section staining , It can be combined with more full slice detection methods , For example, first apply the target detection method to extract the suspected area under low resolution , Then classify and diagnose the focus .

Q： What is the normal sample data source of pathological images ？

A： Because completely healthy human aortic tissue samples are not available , Therefore, the normal sample comes from the normal area of the patient section with mild lesions . For pathological images , Due to the consistency of tissue morphology in a certain range , And the focus area does not affect the normal characterization of other areas , So we can use areas other than lesions as normal samples .

Q： be based on OpenVINO Whether the model optimization process of the tool needs to be completed manually ？

A：OpenVINO Model optimizer （Model Optimizer） It can automatically complete the optimization of the model reasoning process , There is no need for users to manually adjust the model structure .

At the end of the sharing meeting , The lecturer introduced the application process of the industry university cooperation project supported by the Ministry of Education , Encourage more students to join the school enterprise cooperation project , Actively implement and promote laboratory achievements and industrialization .

END