Objective:To establish an automatic classification approach for bone marrow cells based on microscopic hyperspectral imaging and three-dimensional spectral convolutional neural network (Spec-CNN).Methods:The research type is establishment of methodology. The study included 306 newly diagnosed patients' bone marrow smears under Wright's staining from the Department of Hematology of the First Medical Center of the PLA General Hospital from November 1st, 2013 to April 30th, 2024. The high-spectrum data and 4k image data of bone marrow cells were simultaneously collected using a microscopic hyperspectral-4k optical path integrated imaging system (with a spectral resolution of 400—1 000 nm). The high-spectrum data was used for model training, while the 4k image data recognized by morphologists was only used as a reference for labeling the high-spectrum data. The high-spectrum data set was divided into training set, validation set and test set in a ratio of 14∶6∶5. The training set and validation set were used to train and fine-tune the Spec-CNN model, and the test set was used to evaluate the model performance. The sensitivity, specificity ,accuracy ,and Kappa coefficient were calculated for comparing the manual annotation results as gold standard with the intelligent identification results of the Spec-CNN model. Five non-data set samples were used for external validation.Results:The acquired hyperspectral data and 4k imaging dataset comprised of 32 categories and 64 800 bone marrow cells. In the test set, the Spec-CNN model demonstrated weighted-average indicators on classification metrics across 32 cell types: sensitivity 87.79%, specificity 99.31%, and accuracy 98.78%, and Kappa coefficient 0.869. For external validation, the mean correct identification rate of bone marrow cells reached 83.28%.Conclusion:We successfully established an automatic classification method of bone marrow cells based on microscopic hyperspectral imaging and three-dimensional Spec-CNN. This method has a good automatic classification ability for 32 types of bone marrow nucleated cells, which has a certain auxiliary effect on improving the diagnosis efficiency of blood diseases for bone marrow morphologists.

Objective:To establish an automatic classification approach for bone marrow cells based on microscopic hyperspectral imaging and three-dimensional spectral convolutional neural network (Spec-CNN).Methods:The research type is establishment of methodology. The study included 306 newly diagnosed patients' bone marrow smears under Wright's staining from the Department of Hematology of the First Medical Center of the PLA General Hospital from November 1st, 2013 to April 30th, 2024. The high-spectrum data and 4k image data of bone marrow cells were simultaneously collected using a microscopic hyperspectral-4k optical path integrated imaging system (with a spectral resolution of 400—1 000 nm). The high-spectrum data was used for model training, while the 4k image data recognized by morphologists was only used as a reference for labeling the high-spectrum data. The high-spectrum data set was divided into training set, validation set and test set in a ratio of 14∶6∶5. The training set and validation set were used to train and fine-tune the Spec-CNN model, and the test set was used to evaluate the model performance. The sensitivity, specificity ,accuracy ,and Kappa coefficient were calculated for comparing the manual annotation results as gold standard with the intelligent identification results of the Spec-CNN model. Five non-data set samples were used for external validation.Results:The acquired hyperspectral data and 4k imaging dataset comprised of 32 categories and 64 800 bone marrow cells. In the test set, the Spec-CNN model demonstrated weighted-average indicators on classification metrics across 32 cell types: sensitivity 87.79%, specificity 99.31%, and accuracy 98.78%, and Kappa coefficient 0.869. For external validation, the mean correct identification rate of bone marrow cells reached 83.28%.Conclusion:We successfully established an automatic classification method of bone marrow cells based on microscopic hyperspectral imaging and three-dimensional Spec-CNN. This method has a good automatic classification ability for 32 types of bone marrow nucleated cells, which has a certain auxiliary effect on improving the diagnosis efficiency of blood diseases for bone marrow morphologists.

Aiming at the non-stationarity and temporal characteristics of variable-length electrocardiogram(ECG)signals,an arrhythmia identification algorithm is proposed based on continuous wavelet transform and higher-order statistics.Considering the varying number of data points for each sample in variable-length ECG signals,the RR interval interpolation method is employed for data preprocessing,and the signal is decomposed into different time-frequency components using continuous wavelet transform,which enables the network to better extract both temporal and frequency features from the ECG signals.Regarding the issue of insufficient utilization of temporal information,a temporal mining module is introduced based on higher-order statistics and long short-term memory network to capture and learn long-term dependencies in the ECG signals,thereby facilitating the identification and understanding of specific arrhythmia categories.Extensive experiments conducted on the publicly available MIT-BIH ECG database validate the effectiveness and superiority of the proposed method.

Objective:To explore the value of chest low-dose CT (LDCT) in post-discharge follow-up assessments of patients with coronavirus disease 2019 (COVID-19).Methods:The chest CT findings of 58 patients with COVID-19 from March 17 to March 25, 2020 at Remin Hospital of Wuhan University were retrospectively analyzed. Two radiologists independently scored the subjective image quality on a 5-point Likert scale. The signal-to-noise ratio (SNR) and SD air of images and the CT radiation dose parameters were calculated, including the CT volume dose index (CTDI vol), dose length product (DLP), and effective radiation dose ( E). Results:The subjective image quality scores on CT images obtained before and after discharge by readers 1 and 2, were 4.45±0.22, 3.88±0.33 ( P>0.05) and 4.37±0.18, 3.91±0.35 ( P>0.05), respectively. The SNR and SD air in LDCT after discharge were 4.39±0.95 and 7.19±2.41, which were significantly lower than those in routine chest CT before discharge (5.14±1.06, Z=-5.551, P<0.001; 6.48±1.57, Z=-3.217, P<0.001). All of the obtained images were sufficient for diagnosis. The CTDI vol, DLP, and E in LDCT were significantly lower than those in routine CT [(2.41±0.09), (10.53±1.03)mGy, Z=-6.568, P<0.001; (88.03±5.33), (338.74±34.64)mGy·cm, Z=-6.624, P<0.001; and (1.23±0.17), (4.74±0.48)mSv, Z=-5.976, P<0.001]. Conclusions:Patients with COVID-19 can be followed up with low-dose chest CT after discharge.