| Breast cancer is one of the most common malignant tumors in the world,especially in the female population,the incidence of breast cancer is far higher than other cancers.Pathological section is still the gold standard for the diagnosis of breast tumors,but the examination cycle is long and depends on the experience of pathologists,the diagnostic accuracy will be affected by clinical factors.Rapid frozen section is a technique used for pathological diagnosis in breast surgery,but it requires a series of complicated operations and a long waiting time for doctors and patients.Deep learning technology has been widely used in pathological image classification,but current studies are mostly based on images of traditional stained sections,unable to obtain intracellular information and diagnose unstained sections,so it cannot be applied in intraoperative diagnosis.With the rapid development of hyperspectral imaging technology,it has great research potential in medical diagnosis.This technique can obtain the spectral image information of tissue and provide an effective auxiliary examination method for biomedicine.As an alternative to other existing diagnostic techniques,hyperspectral imaging technology offers the advantage of scanning fast,completely non-invasive,non-contact,nonionizing and non-labeling sensing technologies..Therefore,the application of hyperspectral imaging technology to unstained rapid frozen pathological sections of breast tumor tissue can not only reduce the manual reading error,but also shorten the waiting time of doctors and patients during the operation and improve the efficiency of diagnosis and treatment.In this paper,reflection detection mode and fluorescence detection mode of hyperspectral imaging technology were applied to HE staining and unstained rapid freezing pathological sections of breast tumor tissue.Firstly,preprocessing operations such as reflectivity calibration,derivative spectrum and data standardization are carried out on the collected hyperspectral fluorescence data and hyperspectral reflection data,which reduces the interference of environment and system noise and improves the signal-to-noise ratio of hyperspectral data.Then,according to the characteristics of hyperspectral images,a one-dimensional convolution neural network model based on spectral features was built.Finally,the classification accuracy of the reflection hyperspectral data set and fluorescence hyperspectral data set of breast tumor staining sections reached 94.95%and 96.04%respectively.The classification accuracy was 91.85%and 92.06%on the reflectance and fluorescence hyperspectral datasets of unstained slices,respectively.Finally,according to the characteristics of complex texture structure and rich detail information of pathological sections,combined with the rich multi-dimensional information of hyperspectral data,a 3D convolution neural network(3D-CNN)classification method combining spatial structure and spectral information was proposed,which effectively improved the classification accuracy of stained and unstained sections.Through the comparison of different samples,different image acquisition modes and different classification methods,the feasibility of the application of unstained rapid frozen pathological sections in rapid intraoperative diagnosis was finally verified.In this paper,we study shows that compared with the traditional image classification research,using set highlights like technology,not only for dyeing biopsy can obtain the information inside the cell,improve the effect of classification,the dyeing slice can not achieve good classification effect,which makes not biopsy for pathological dyeing auxiliary accurate diagnosis possible,of intraoperative rapid diagnosis equipment research and development provides a theoretical basis,It is expected to shorten the waiting time of doctors and patients for breast cancer surgery and provide a new method for rapid pathological diagnosis of breast cancer during operation. |
PDF Full Text Request