Background With the gradual increase of infertility in the world,among which male sperm problems are the main factor for infertility,more and more couples are using computer-assisted sperm analysis(CASA)to assist in t...Background With the gradual increase of infertility in the world,among which male sperm problems are the main factor for infertility,more and more couples are using computer-assisted sperm analysis(CASA)to assist in the analysis and treatment of infertility.Meanwhile,the rapid development of deep learning(DL)has led to strong results in image classification tasks.However,the classification of sperm images has not been well studied in current deep learning methods,and the sperm images are often affected by noise in practical CASA applications.The purpose of this article is to investigate the anti-noise robustness of deep learning classification methods applied on sperm images.Methods The SVIA dataset is a publicly available large-scale sperm dataset containing three subsets.In this work,we used subset-C,which provides more than 125,000 independent images of sperms and impurities,including 121,401 sperm images and 4,479 impurity images.To investigate the anti-noise robustness of deep learning classification methods applied on sperm images,we conducted a comprehensive comparative study of sperm images using many convolutional neural network(CNN)and visual transformer(VT)deep learning methods to find the deep learning model with the most stable anti-noise robustness.Results This study proved that VT had strong robustness for the classification of tiny object(sperm and impurity)image datasets under some types of conventional noise and some adversarial attacks.In particular,under the influence of Poisson noise,accuracy changed from 91.45%to 91.08%,impurity precison changed from 92.7%to 91.3%,impurity recall changed from 88.8%to 89.5%,and impurity F1-score changed 90.7%to 90.4%.Meanwhile,sperm precision changed from 90.9%to 90.5%,sperm recall changed from 92.5%to 93.8%,and sperm F1-score changed from 92.1%to 90.4%.Conclusion Sperm image classification may be strongly affected by noise in current deep learning methods;the robustness with regard to noise of VT methods based on global information is greater than that of CNN methods based on local information,indicating that the robustness with regard to noise is reflected mainly in global information.展开更多
Microsatellites have recently opened windows of frequent and low cost missions for planetary exploration. The performance of gamma-ray and neutron spectrometers on future microsatellite missions is simulated to assess...Microsatellites have recently opened windows of frequent and low cost missions for planetary exploration. The performance of gamma-ray and neutron spectrometers on future microsatellite missions is simulated to assess the possibility of observation of hydrogen and major elements, given their concentration on the observation target. The measured elemental abundance will provide important geological constraints, and some of them may serve as space resources. Four different types of target bodies with various hydrogen concentrations in the range of 0 - 20,000 ppm are assumed as target compositions;Earth’s core, C-type, S-type and Martian meteorites. Gamma-ray and neutron emission rates show unique footprints that are related to the different elemental compositions. The starting point is the solid angle subtended between observation target and spectrometers that allow estimating the gamma-ray and neutron count rates emitted by the celestial bodies. In this work, three types of gamma-ray detectors;high-purity germanium (HPGe), CeBr3 and LaBr3(Ce), a neutron spectrometer combining a lithium glass scintillator with a boron loaded plastic scintillator and a dual mode spectrometer Cs2LiYCl6(Ce) (CLYC) are simulated, focusing on their observation backgrounds as a model case for microsatellite based measurements. The background count level of both gamma-ray (except for the LaBr3 detector) and neutron count rates was negligible under these particular conditions. The gamma-ray detectors were compared by the figure of merit, which was determined by their efficiency and energy resolution. It was found that each detector has unique advantages. The HPGe detector has the highest figure of merit due to its excellent energy resolution, whereas the CLYC detector is low in weight and power consumption due to its dual sensitivity to gamma-ray and neutron. The CeBr3 detector is an intermediate choice. The neutron count rates are calculated separately in three energy ranges, i.e. , thermal (<0.5 eV), epithermal (0.5 eV - 500 keV), and fast (>500 keV), as a function of the hydrogen concentration in the 0 - 20,000 ppm range. The thermal and epithermal neutron count rates are found to decrease with hydrogen concentration, while the fast neutron count rate increases with the target average atomic mass. The optimal detector should be decided by the mission restraints on mass, power consumption, and heat thermal design.展开更多
Background Histopathology diagnosis is often regarded as the final diagnostic method for malignant tumors;however,it has some drawbacks.This study explored a computer-aided diagnostic method that can be used to identi...Background Histopathology diagnosis is often regarded as the final diagnostic method for malignant tumors;however,it has some drawbacks.This study explored a computer-aided diagnostic method that can be used to identify benign and malignant gastric cancer using histopathological images.Methods The most suitable process was selected through multiple experiments by comparing multiple meth-ods and features for classification.First,the U-net was applied to segment the image.Next,the nucleus was extracted from the segmented image,and the minimum spanning tree(MST)diagram structure that can cap-ture the topological information was drawn.The third step was to extract the graph-curvature features of the histopathological image according to the MST image.Finally,by inputting the graph-curvature features into the classifier,the recognition results for benign or malignant cancer can be obtained.Results During the experiment,we used various methods for comparison.In the image segmentation stage,U-net,watershed algorithm,and Otsu threshold segmentation methods were used.We found that the U-net method,combined with multiple indicators,was the most suitable for segmentation of histopathological images.In the feature extraction stage,in addition to extracting graph-edge and graph-curvature features,several basic im-age features were extracted,including the red,green and blue feature,gray-level co-occurrence matrix feature,histogram of oriented gradient feature,and local binary pattern feature.In the classifier design stage,we exper-imented with various methods,such as support vector machine(SVM),random forest,artificial neural network,K nearest neighbors,VGG-16,and inception-V3.Through comparison and analysis,it was found that classifica-tion results with an accuracy of 98.57%can be obtained by inputting the graph-curvature feature into the SVM classifier.展开更多
Background Breast cancer has the highest prevalence among all cancers in women globally.The classification of histopathological images in the diagnosis of breast cancers is an area of clinical concern.In computer-aide...Background Breast cancer has the highest prevalence among all cancers in women globally.The classification of histopathological images in the diagnosis of breast cancers is an area of clinical concern.In computer-aided diagnosis,most traditional classification models use a single network to extract features,although this approach has significant limitations.Moreover,many networks are trained and optimized on patient-level datasets,ignoring lower-level data labels.Methods This paper proposed a deep ensemble model based on image-level labels for the binary classification of breast histopathological images of benign and malignant lesions.First,the BreaKHis dataset was randomly divided into training,validation,and test sets.Then,data augmentation techniques were used to balance the numbers of benign and malignant samples.Third,based on their transfer learning performance and the complementarity between networks,VGG16,Xception,ResNet50,and DenseNet201 were selected as base classifiers.Results In a ensemble network model with accuracy as the weight,the image-level binary classification achieved an accuracy of 98.90%.To verify the capabilities of our method,it was experimentally compared with the latest transformer and multilayer perception(MLP)models on the same dataset.Our ensemble model showed a 5%-20%advantage,emphasizing its far-reaching abilities in classification tasks.Conclusions This research focuses on improving the performance of a classification model with an ensemble algorithm.Transfer learning has an essential role in classification of small datasets,improving training speed and accuracy.Our model may outperform many existing approaches with respect to accuracy and has applications in the field of auxiliary medical diagnosis.展开更多
基金supported by the National Natural Science Foundation of China(Grant No.82220108007).
文摘Background With the gradual increase of infertility in the world,among which male sperm problems are the main factor for infertility,more and more couples are using computer-assisted sperm analysis(CASA)to assist in the analysis and treatment of infertility.Meanwhile,the rapid development of deep learning(DL)has led to strong results in image classification tasks.However,the classification of sperm images has not been well studied in current deep learning methods,and the sperm images are often affected by noise in practical CASA applications.The purpose of this article is to investigate the anti-noise robustness of deep learning classification methods applied on sperm images.Methods The SVIA dataset is a publicly available large-scale sperm dataset containing three subsets.In this work,we used subset-C,which provides more than 125,000 independent images of sperms and impurities,including 121,401 sperm images and 4,479 impurity images.To investigate the anti-noise robustness of deep learning classification methods applied on sperm images,we conducted a comprehensive comparative study of sperm images using many convolutional neural network(CNN)and visual transformer(VT)deep learning methods to find the deep learning model with the most stable anti-noise robustness.Results This study proved that VT had strong robustness for the classification of tiny object(sperm and impurity)image datasets under some types of conventional noise and some adversarial attacks.In particular,under the influence of Poisson noise,accuracy changed from 91.45%to 91.08%,impurity precison changed from 92.7%to 91.3%,impurity recall changed from 88.8%to 89.5%,and impurity F1-score changed 90.7%to 90.4%.Meanwhile,sperm precision changed from 90.9%to 90.5%,sperm recall changed from 92.5%to 93.8%,and sperm F1-score changed from 92.1%to 90.4%.Conclusion Sperm image classification may be strongly affected by noise in current deep learning methods;the robustness with regard to noise of VT methods based on global information is greater than that of CNN methods based on local information,indicating that the robustness with regard to noise is reflected mainly in global information.
基金This paper is a part of the outcome research performed under a Waseda University Grant for Special Research Project(Project number:2017B-208).
文摘Microsatellites have recently opened windows of frequent and low cost missions for planetary exploration. The performance of gamma-ray and neutron spectrometers on future microsatellite missions is simulated to assess the possibility of observation of hydrogen and major elements, given their concentration on the observation target. The measured elemental abundance will provide important geological constraints, and some of them may serve as space resources. Four different types of target bodies with various hydrogen concentrations in the range of 0 - 20,000 ppm are assumed as target compositions;Earth’s core, C-type, S-type and Martian meteorites. Gamma-ray and neutron emission rates show unique footprints that are related to the different elemental compositions. The starting point is the solid angle subtended between observation target and spectrometers that allow estimating the gamma-ray and neutron count rates emitted by the celestial bodies. In this work, three types of gamma-ray detectors;high-purity germanium (HPGe), CeBr3 and LaBr3(Ce), a neutron spectrometer combining a lithium glass scintillator with a boron loaded plastic scintillator and a dual mode spectrometer Cs2LiYCl6(Ce) (CLYC) are simulated, focusing on their observation backgrounds as a model case for microsatellite based measurements. The background count level of both gamma-ray (except for the LaBr3 detector) and neutron count rates was negligible under these particular conditions. The gamma-ray detectors were compared by the figure of merit, which was determined by their efficiency and energy resolution. It was found that each detector has unique advantages. The HPGe detector has the highest figure of merit due to its excellent energy resolution, whereas the CLYC detector is low in weight and power consumption due to its dual sensitivity to gamma-ray and neutron. The CeBr3 detector is an intermediate choice. The neutron count rates are calculated separately in three energy ranges, i.e. , thermal (<0.5 eV), epithermal (0.5 eV - 500 keV), and fast (>500 keV), as a function of the hydrogen concentration in the 0 - 20,000 ppm range. The thermal and epithermal neutron count rates are found to decrease with hydrogen concentration, while the fast neutron count rate increases with the target average atomic mass. The optimal detector should be decided by the mission restraints on mass, power consumption, and heat thermal design.
基金supported by the National Natural Science Foundation of China(Grant No.82220108007).
文摘Background Histopathology diagnosis is often regarded as the final diagnostic method for malignant tumors;however,it has some drawbacks.This study explored a computer-aided diagnostic method that can be used to identify benign and malignant gastric cancer using histopathological images.Methods The most suitable process was selected through multiple experiments by comparing multiple meth-ods and features for classification.First,the U-net was applied to segment the image.Next,the nucleus was extracted from the segmented image,and the minimum spanning tree(MST)diagram structure that can cap-ture the topological information was drawn.The third step was to extract the graph-curvature features of the histopathological image according to the MST image.Finally,by inputting the graph-curvature features into the classifier,the recognition results for benign or malignant cancer can be obtained.Results During the experiment,we used various methods for comparison.In the image segmentation stage,U-net,watershed algorithm,and Otsu threshold segmentation methods were used.We found that the U-net method,combined with multiple indicators,was the most suitable for segmentation of histopathological images.In the feature extraction stage,in addition to extracting graph-edge and graph-curvature features,several basic im-age features were extracted,including the red,green and blue feature,gray-level co-occurrence matrix feature,histogram of oriented gradient feature,and local binary pattern feature.In the classifier design stage,we exper-imented with various methods,such as support vector machine(SVM),random forest,artificial neural network,K nearest neighbors,VGG-16,and inception-V3.Through comparison and analysis,it was found that classifica-tion results with an accuracy of 98.57%can be obtained by inputting the graph-curvature feature into the SVM classifier.
基金supported by the National Natural Science Foundation of China(Grant No.61806047).
文摘Background Breast cancer has the highest prevalence among all cancers in women globally.The classification of histopathological images in the diagnosis of breast cancers is an area of clinical concern.In computer-aided diagnosis,most traditional classification models use a single network to extract features,although this approach has significant limitations.Moreover,many networks are trained and optimized on patient-level datasets,ignoring lower-level data labels.Methods This paper proposed a deep ensemble model based on image-level labels for the binary classification of breast histopathological images of benign and malignant lesions.First,the BreaKHis dataset was randomly divided into training,validation,and test sets.Then,data augmentation techniques were used to balance the numbers of benign and malignant samples.Third,based on their transfer learning performance and the complementarity between networks,VGG16,Xception,ResNet50,and DenseNet201 were selected as base classifiers.Results In a ensemble network model with accuracy as the weight,the image-level binary classification achieved an accuracy of 98.90%.To verify the capabilities of our method,it was experimentally compared with the latest transformer and multilayer perception(MLP)models on the same dataset.Our ensemble model showed a 5%-20%advantage,emphasizing its far-reaching abilities in classification tasks.Conclusions This research focuses on improving the performance of a classification model with an ensemble algorithm.Transfer learning has an essential role in classification of small datasets,improving training speed and accuracy.Our model may outperform many existing approaches with respect to accuracy and has applications in the field of auxiliary medical diagnosis.
