Sex Estimation of Medial Aspect of the Ischiopubic Ramus in Adults Based on Deep Learning / 法医学杂志

OBJECTIVES@#To investigate the reliability and accuracy of deep learning technology in automatic sex estimation using the 3D reconstructed images of the computed tomography (CT) from the Chinese Han population.@*METHODS@#The pelvic CT images of 700 individuals (350 males and 350 females) of the Chinese Han population aged 20 to 85 years were collected and reconstructed into 3D virtual skeletal models. The feature region images of the medial aspect of the ischiopubic ramus (MIPR) were intercepted. The Inception v4 was adopted as the image recognition model, and two methods of initial learning and transfer learning were used for training. Eighty percent of the individuals' images were randomly selected as the training and validation dataset, and the remaining were used as the test dataset. The left and right sides of the MIPR images were trained separately and combinedly. Subsequently, the models' performance was evaluated by overall accuracy, female accuracy, male accuracy, etc.@*RESULTS@#When both sides of the MIPR images were trained separately with initial learning, the overall accuracy of the right model was 95.7%, the female accuracy and male accuracy were both 95.7%; the overall accuracy of the left model was 92.1%, the female accuracy was 88.6% and the male accuracy was 95.7%. When the left and right MIPR images were combined to train with initial learning, the overall accuracy of the model was 94.6%, the female accuracy was 92.1% and the male accuracy was 97.1%. When the left and right MIPR images were combined to train with transfer learning, the model achieved an overall accuracy of 95.7%, and the female and male accuracies were both 95.7%.@*CONCLUSIONS@#The use of deep learning model of Inception v4 and transfer learning algorithm to construct a sex estimation model for pelvic MIPR images of Chinese Han population has high accuracy and well generalizability in human remains, which can effectively estimate the sex in adults.

Automatic sleep staging algorithm for stochastic depth residual networks based on transfer learning / 生物医学工程学杂志

The existing automatic sleep staging algorithms have the problems of too many model parameters and long training time, which in turn results in poor sleep staging efficiency. Using a single channel electroencephalogram (EEG) signal, this paper proposed an automatic sleep staging algorithm for stochastic depth residual networks based on transfer learning (TL-SDResNet). Firstly, a total of 30 single-channel (Fpz-Cz) EEG signals from 16 individuals were selected, and after preserving the effective sleep segments, the raw EEG signals were pre-processed using Butterworth filter and continuous wavelet transform to obtain two-dimensional images containing its time-frequency joint features as the input data for the staging model. Then, a ResNet50 pre-trained model trained on a publicly available dataset, the sleep database extension stored in European data format (Sleep-EDFx) was constructed, using a stochastic depth strategy and modifying the output layer to optimize the model structure. Finally, transfer learning was applied to the human sleep process throughout the night. The algorithm in this paper achieved a model staging accuracy of 87.95% after conducting several experiments. Experiments show that TL-SDResNet50 can accomplish fast training of a small amount of EEG data, and the overall effect is better than other staging algorithms and classical algorithms in recent years, which has certain practical value.

Automated Classification of Inherited Retinal Diseases in Optical Coherence Tomography Images Using Few-shot Learning / 生物医学与环境科学（英文）

OBJECTIVE@#To develop a few-shot learning (FSL) approach for classifying optical coherence tomography (OCT) images in patients with inherited retinal disorders (IRDs).@*METHODS@#In this study, an FSL model based on a student-teacher learning framework was designed to classify images. 2,317 images from 189 participants were included. Of these, 1,126 images revealed IRDs, 533 were normal samples, and 658 were control samples.@*RESULTS@#The FSL model achieved a total accuracy of 0.974-0.983, total sensitivity of 0.934-0.957, total specificity of 0.984-0.990, and total F1 score of 0.935-0.957, which were superior to the total accuracy of the baseline model of 0.943-0.954, total sensitivity of 0.866-0.886, total specificity of 0.962-0.971, and total F1 score of 0.859-0.885. The performance of most subclassifications also exhibited advantages. Moreover, the FSL model had a higher area under curves (AUC) of the receiver operating characteristic (ROC) curves in most subclassifications.@*CONCLUSION@#This study demonstrates the effective use of the FSL model for the classification of OCT images from patients with IRDs, normal, and control participants with a smaller volume of data. The general principle and similar network architectures can also be applied to other retinal diseases with a low prevalence.

Retinal Lesion Segmentation Using Transfer Learning with an Encoder-Decoder CNN / Segmentación de Lesiones en la Retina Usando Transferencia de Conocimiento con CNN Encoder-Decoder

ABSTRACT Deep learning (DL) techniques achieve high performance in the detection of illnesses in retina images, but the majority of models are trained with different databases for solving one specific task. Consequently, there are currently no solutions that can be used for the detection/segmentation of a variety of illnesses in the retina in a single model. This research uses Transfer Learning (TL) to take advantage of previous knowledge generated during model training of illness detection to segment lesions with encoder-decoder Convolutional Neural Networks (CNN), where the encoders are classical models like VGG-16 and ResNet50 or variants with attention modules. This shows that it is possible to use a general methodology using a single fundus image database for the detection/segmentation of a variety of retinal diseases achieving state-of-the-art results. This model could be in practice more valuable since it can be trained with a more realistic database containing a broad spectrum of diseases to detect/segment illnesses without sacrificing performance. TL can help achieve fast convergence if the samples in the main task (Classification) and sub-tasks (Segmentation) are similar. If this requirement is not fulfilled, the parameters start from scratch.

RESUMEN Las técnicas de Deep Learning (DL) han demostrado un buen desempeño en la detección de anomalías en imágenes de retina, pero la mayoría de los modelos son entrenados en diferentes bases de datos para resolver una tarea en específico. Como consecuencia, actualmente no se cuenta con modelos que se puedan usar para la detección/segmentación de varias lesiones o anomalías con un solo modelo. En este artículo, se utiliza Transfer Learning (TL) con la cual se aprovecha el conocimiento adquirido para determinar si una imagen de retina tiene o no una lesión. Con este conocimiento se segmenta la imagen utilizando una red neuronal convolucional (CNN), donde los encoders o extractores de características son modelos clásicos como VGG-16 y ResNet50 o variantes con módulos de atención. Se demuestra así, que es posible utilizar una metodología general con bases de datos de retina para la detección/ segmentación de lesiones en la retina alcanzando resultados como los que se muestran en el estado del arte. Este modelo puede ser entrenado con bases de datos más reales que contengan una gama de enfermedades para detectar/ segmentar sin sacrificar rendimiento. TL puede ayudar a conseguir una convergencia rápida del modelo si la base de datos principal (Clasificación) se parece a la base de datos de las tareas secundarias (Segmentación), si esto no se cumple los parámetros básicamente comienzan a ajustarse desde cero.

Diagnosis model for fusarium identification of fungal keratitis based on transfer learning and data augmentation / 国际眼科杂志(Guoji Yanke Zazhi)

@#AIM: To construct and evaluate a diagnostic model based on transfer learning and data augmentation as a non-invasive tool for fusarium identification of fungal keratitis. <p>METHODS: A retrospective study. In this study, 2 157 images of fungal keratitis patients who underwent <i>in vivo</i> confocal microscopy examination in the Department of Ophthalmology of the people's Hospital of Guangxi Zhuang Autonomous Region from March 2017 to January 2020 were included as the dataset, which was classified according to the results of microbial culture. The dataset was subsequently randomly divided into training set(1 380 images), validation set(345 images)and test set(432 images). We used the transfer learning Inception-ResNet V2 network to construct a diagnostic model, and to compare the performance of the model trained on different datasets. The performance of the diagnostic model evaluated with specificity, sensitivity, accuracy, and area under the receiver operating characteristics curve(AUC).<p>RESULTS: The model trained with the original training set had a specificity rate of 71.6%, a sensitivity rate of 72.0%, an accuracy rate of 71.8% and AUC of 0.785(95%<i>CI</i>: 0.742-0.828, <i>P</i><0.0001). And the model trained with the augmented training set had a specificity rate of 76.6%, a sensitivity rate of 83.1%, an accuracy rate of 79.9% and AUC of 0.876(95%<i>CI</i>: 0.843-0.909, <i>P</i><0.0001), which made the model's prediction performance boost.<p>CONCLUSION: In this study, we constructed an intelligent diagnosis system for fungal keratitis fusarium through transfer learning, which has higher accuracy, and realized the intelligent diagnosis of fungal keratitis pathogen fusarium. Furthermore, we verified that the data augmentation strategy can improve the performance of the intelligent diagnosis system when the original dataset is limited, and this method can be used for intelligent diagnosis and identification of fungal keratitis pathogen fusarium.

Research on the automatic classification system of pterygium based on deep learning / 国际眼科杂志(Guoji Yanke Zazhi)

@#AIM: To evaluate the application value of the automatic classification and diagnosis system of pterygium based on deep learning.<p>METHODS: A total of 750 images of normal, observational and operative anterior sections of pterygium were collected from the Affiliated Eye Hospital of Nanjing Medical University between May 2020 and April 2021. Seven triclassification models were respectively trained with original data set and enhanced data set. Totally 470 clinical images were tested, and the generalization ability of the model before and after data enhancement was compared to determine the best model for the automatic classification system of pterygium.<p>RESULTS:The average sensitivity, specificity and AUC of the best model trained on the original data set were 92.55%, 96.86% and 94.70% respectively. After data was enhanced, the sensitivity, specificity and AUC of different models were increased by 3.7%, 1.9% and 2.7% on average. The sensitivity, specificity and AUC of the EfficientNetB7 model trained on the enhanced data set were 93.63%, 97.34% and 95.47% respectively.<p>CONDLUSION: The EfficientNetB7 model, which was trained on the enhanced data set, achieves the best classification result and can be used in the automatic classification system of pterygium.This automatic classification system can diagnose diseases about pterygium better, and it is expected to be an effective screening tool for primary medical care. It also provides reference for the research on the refinement and grading of pterygium.

Clasificación de Imágenes de Neumonía a causa de Covid-19 utilizando Transfer-Learning basado en Redes Convolucionales / Classification of Images of Pneumonia Due to Covid-19 Using TransferLearning, Based on Convolutional Networks

La Inteligencia Artificial ha ayudado a lidiar diferentes problemas relacionados con los datos masivos y a su vez con su tratamiento, diagnóstico y detección de enfermedades como la que actualmente nos preocupa, la Covid-19. El objetivo de esta investigación ha sido analizar y desarrollar la clasificación de imágenes de neumonía a causa de covid-19 para un diagnostico efectivo y óptimo. Se ha usado Transfer-Learning aplicando ResNet, DenseNet, Poling y Dense layer para la elaboración de los modelos de red propios Covid-UPeU y Covid-UPeU-TL, utilizando las plataformas Kaggle y Google colab, donde se realizaron 4 experimentos. El resultado con una mejor clasificación de imágenes se obtuvo en el experimento 4 prueba N°2 con el modelo Covid-UPeU-TL donde Acc.Train: 0.9664 y Acc.Test: 0.9851. Los modelos implementados han sido desarrollados con el propósito de tener una visión holística de los factores para la optimización en la clasificación de imágenes de neumonía a causa de COVID-19(AU)

Artificial Intelligence has helped to deal with different problems related to massive data in turn to the treatment, diagnosis and detection of diseases such as the one that currently has us in concern, Covid-19. The objective of this research has been to analyze and develop the classification of images of pneumonia due to covid-19 for an effective and optimal diagnosis. Transfer-Learning has been used applying ResNet, DenseNet, Poling and Dense layer for the elaboration of the own network models Covid-Upeu and Covid-UpeU-TL, using Kaggle and Google colab platforms, where 4 experiments have been carried out. The result with a better classification of images was obtained in experiment 4 test N ° 2 with the Covid-UPeU-TL model where Acc.Train: 0.9664 and Acc.Test: 0.9851. The implemented models have been developed with the purpose of having a holistic view of the factors for optimization in the classification of COVID-19 images(AU)

Epilepsy detection and analysis method for specific patient based on data augmentation and deep learning / 生物医学工程学杂志

In recent years, epileptic seizure detection based on electroencephalogram (EEG) has attracted the widespread attention of the academic. However, it is difficult to collect data from epileptic seizure, and it is easy to cause over fitting phenomenon under the condition of few training data. In order to solve this problem, this paper took the CHB-MIT epilepsy EEG dataset from Boston Children's Hospital as the research object, and applied wavelet transform for data augmentation by setting different wavelet transform scale factors. In addition, by combining deep learning, ensemble learning, transfer learning and other methods, an epilepsy detection method with high accuracy for specific epilepsy patients was proposed under the condition of insufficient learning samples. In test, the wavelet transform scale factors 2, 4 and 8 were set for experimental comparison and verification. When the wavelet scale factor was 8, the average accuracy, average sensitivity and average specificity was 95.47%, 93.89% and 96.48%, respectively. Through comparative experiments with recent relevant literatures, the advantages of the proposed method were verified. Our results might provide reference for the clinical application of epilepsy detection.

Parameter transfer learning based on shallow visual geometry group network and its application in motor imagery classification / 生物医学工程学杂志

Transfer learning is provided with potential research value and application prospect in motor imagery electroencephalography (MI-EEG)-based brain-computer interface (BCI) rehabilitation system, and the source domain classification model and transfer strategy are the two important aspects that directly affect the performance and transfer efficiency of the target domain model. Therefore, we propose a parameter transfer learning method based on shallow visual geometry group network (PTL-sVGG). First, Pearson correlation coefficient is used to screen the subjects of the source domain, and the short-time Fourier transform is performed on the MI-EEG data of each selected subject to acquire the time-frequency spectrogram images (TFSI). Then, the architecture of VGG-16 is simplified and the block design is carried out, and the modified sVGG model is pre-trained with TFSI of source domain. Furthermore, a block-based frozen-fine-tuning transfer strategy is designed to quickly find and freeze the block with the greatest contribution to sVGG model, and the remaining blocks are fine-tuned by using TFSI of target subjects to obtain the target domain classification model. Extensive experiments are conducted based on public MI-EEG datasets, the average recognition rate and Kappa value of PTL-sVGG are 94.9% and 0.898, respectively. The results show that the subjects' optimization is beneficial to improve the model performance in source domain, and the block-based transfer strategy can enhance the transfer efficiency, realizing the rapid and effective transfer of model parameters across subjects on the datasets with different number of channels. It is beneficial to reduce the calibration time of BCI system, which promote the application of BCI technology in rehabilitation engineering.

Cross-subject electroencephalogram emotion recognition based on maximum classifier discrepancy / 生物医学工程学杂志

Affective brain-computer interfaces (aBCIs) has important application value in the field of human-computer interaction. Electroencephalogram (EEG) has been widely concerned in the field of emotion recognition due to its advantages in time resolution, reliability and accuracy. However, the non-stationary characteristics and individual differences of EEG limit the generalization of emotion recognition model in different time and different subjects. In this paper, in order to realize the recognition of emotional states across different subjects and sessions, we proposed a new domain adaptation method, the maximum classifier difference for domain adversarial neural networks (MCD_DA). By establishing a neural network emotion recognition model, the shallow feature extractor was used to resist the domain classifier and the emotion classifier, respectively, so that the feature extractor could produce domain invariant expression, and train the decision boundary of classifier learning task specificity while realizing approximate joint distribution adaptation. The experimental results showed that the average classification accuracy of this method was 88.33% compared with 58.23% of the traditional general classifier. It improves the generalization ability of emotion brain-computer interface in practical application, and provides a new method for aBCIs to be used in practice.

DME-Deep: A Computerize Tool for Detection of Diabetic Macular Edema Grading Based on Multilayer Deep Learning and Transfer Learning

Diabetic macular edema (DME) is a common disease of diabetic retinopathy (DR). Due to the infection of DME disease, many patients’ vision is lost. To cure DME eye disease, early detection and treatment are very important and vital steps. To automatically diagnosis DEM disease, several studies were developed by detection of the macula center which is dependent on optic disc (OD) location. In this paper, a novel features pre-training based model was proposed based on dense convolutional neural network (DCNN) to diagnose DME related disease. As a result, a computerize tool “DME-Deep” for detection of DME-based grading system was implemented through a new dense deep learning model and feature’s transfer learning approaches. This DCNN model was developed by adding new five convolutional and one dropout layers to the network. The DME-Deep system was tested on three different datasets, which obtained from online sources. To train the DCNN model for features learning, the 1650 retinal fundus images were utilized from the Hamilton HEI-MED, ISBI 2018 IDRiD and MESSIDOR datasets. On datasets, the DME-Deep achieved 91.2% of accuracy, 87.5% of sensitivity and 94.4% of specificity. Compare to obtain hand-crafted features, the automatic feature’ learning it provided favorable results. Hence, the experimental results also indicate that this DME-Deep system can automatically assist ophthalmologists in finding DEM eye-related disease.

An Evaluation of Lower Limb Motor Ability Based on Xception-LSTM / 中国康复理论与实践

Objective:To establish an algorithm to quantitatively evaluate the lower limb motor ability. Methods:From August, 2016 to March, 2017, 40 subjects were divided into young healthy group (n = 20), middle-aged group (n = 10) and elderly group (n = 10). The gait video, knee angle and ground reaction force of the subjects were collected, and the gait contour was extracted from the gait video by using the ViBe algorithm. The gait image feature was extracted by Xception-LSTM, and fused it with the knee joint angle and the ground reaction force in the feature layer. The fusion features were reduced in dimension by kernel principal component analysis, and the gait ability score (GAS) was established. All the subjects were assessed with Wisconsin Gait Scale (WGS). Results:GAS was less in middle-aged group and elderly group than in the young healthy group (t > 4.164, P < 0.01), and was less in the elderly group than in the middle-aged group (t = 7.338, P < 0.01). GAS was negative correlated with the score of WGS (r = -0.91, P < 0.01). Conclusion:The lower limb exercise ability could be quantified with GAS, which may be applied in developing rehabilitation and fitting walking aids.

Prognostic model of small sample critical diseases based on transfer learning / 生物医学工程学杂志

Aiming at the problem that the small samples of critical disease in clinic may lead to prognostic models with poor performance of overfitting, large prediction error and instability, the long short-term memory transferring algorithm (transLSTM) was proposed. Based on the idea of transfer learning, the algorithm leverages the correlation between diseases to transfer information of different disease prognostic models, constructs the effictive model of target disease of small samples with the aid of large data of related diseases, hence improves the prediction performance and reduces the requirement for target training sample quantity. The transLSTM algorithm firstly uses the related disease samples to pretrain partial model parameters, and then further adjusts the whole network with the target training samples. The testing results on MIMIC-Ⅲ database showed that compared with traditional LSTM classification algorithm, the transLSTM algorithm had 0.02-0.07 higher AUROC and 0.05-0.14 larger AUPRC, while its number of training iterations was only 39%-64% of the traditional algorithm. The results of application on sepsis revealed that the transLSTM model of only 100 training samples had comparable mortality prediction performance to the traditional model of 250 training samples. In small sample situations, the transLSTM algorithm has significant advantages with higher prediciton accuracy and faster training speed. It realizes the application of transfer learning in the prognostic model of critical disease with small samples.

A TrAdaBoost-based method for detecting multiple subjects' P300 potentials / 生物医学工程学杂志

Individual differences of P300 potentials lead to that a large amount of training data must be collected to construct pattern recognition models in P300-based brain-computer interface system, which may cause subjects' fatigue and degrade the system performance. TrAdaBoost is a method that transfers the knowledge from source area to target area, which improves learning effect in the target area. Our research purposed a TrAdaBoost-based linear discriminant analysis and a TrAdaBoost-based support vector machine to recognize the P300 potentials across multiple subjects. This method first trains two kinds of classifiers separately by using the data deriving from a small amount of data from same subject and a large amount of data from different subjects. Then it combines all the classifiers with different weights. Compared with traditional training methods that use only a small amount of data from same subject or mixed different subjects' data to directly train, our algorithm improved the accuracies by 19.56% and 22.25% respectively, and improved the information transfer rate of 14.69 bits/min and 15.76 bits/min respectively. The results indicate that the TrAdaBoost-based method has the potential to enhance the generalization ability of brain-computer interface on the individual differences.

Construction of gastroscopic image recognition model based on transfer learning and its application in gastric cancer diagnosis / 第二军医大学学报

Objective To construct a gastroscopic image recognition model based on transfer learning and to explore its diagnostic value for gastric cancer. Methods The clear white-light gastroscopic images from 2 001 gastric cancer patients, 2 119 gastric ulcer patients and 2 168 chronic gastritis patients were collected. All these images were divided into training set image group (1 851 gastric cancer, 1 969 gastric ulcer, and 2 018 chronic gastritis) and testing set image group (150 gastric cancer, 150 gastric ulcer, and 150 chronic gastritis). Champion models VGG19, ResNet50 and Inception-V3 in ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) competition were used as pre-trained models. These models were revised for model training. The training set images were assigned to train the above 3 models, and the testing set images were assigned to validate the models. The whole training process was divided into 2 steps (pre-training and fine-tuning). Results It was found that ResNet50 ranked No.1 in terms of testing accuracy. Its diagnostic accuracy for gastric cancer, gastric ulcer and chronic gastritis reached 93%, 92% and 88%, respectively. Conclusion Based on transfer learning, the gastroscopic image recognition software model constructed by ResNet50 model can more accurately differentiate gastric cancer from benign gastric diseases (gastric ulcer and chronic gastritis).

Classification and diagnosis of ultrasound images with breast tumors based on transfer learning / 中国医学影像技术

Objective: To investigate the value of transfer learning methods in classification of ultrasound images of benign and malignant breast tumors. Methods Ultrasonic features of histopathologically proved breast tumors in 447 patients were retrospectively analyzed. The features of original images were extracted using the method of principal component analysis. Matlab 7.0 software was used for achieving transfer learning method. Finally, the quantitative image characteristics were inputted into the program in order to use new methods of transfer learning for identifying the benign and malignant breast tumors. Results The quantitative parameters of ultrasound images with malignant breast tumors, such as edge roughness, firmness, neighborhood gray-tone difference matrix roughness, echo difference between the posterior and peripheral areas of the masses, and the horizontal high-frequency and vertical low-frequency components-histogram energy were significantly higher than those of the benign breast tumors (all P<0.05). The sensitivity, specificity, the accuracy of the ultrasound and transfer learning method in diagnosis of malignant breast tumors was 96.21% (127/132) and 96.04% (97/101), 66.35% (209/315) and 98.49% (196/199), 75.17% (336/447) and 97.67% (293/300), respectively. Conclusion Quantitative ultrasonic features can provide objective quantitative parameters for identification of benign and malignant breast tumors. Transfer learning methods can effectively classify ultrasound images with benign and malignant breast tumors.

Diabetic retinopathy detection algorithm based on transfer learning / 中华实验眼科杂志

Objective To investigate a diabetic retinopathy ( DR ) detection algorithm based on transfer learning in small sample dataset. Methods Total of 4465 fundus color photographs taken by Gaoyao People ' s Hospital was used as the full dataset. The model training strategies using fixed pre-trained parameters and fine-tuning pre-trained parameters were used as the transfer learning group to compare with the non-transfer learning strategy that randomly initializes parameters. These three training strategies were applied to the training of three deep learning networks:ResNet50,Inception V3 and NASNet. In addition,a small dataset randomly extracted from the full dataset was used to study the impact of the reduction of training data on different strategies. The accuracy and training time of the diagnostic model were used to analyze the performance of different training strategies. Results The best results in different network architectures were chosen. The accuracy of the model obtained by fine-tuning pre-training parameters strategy was 90. 9％,which was higher than the strategy of fixed pre-training parameters (88. 1％) and the strategy of randomly initializing parameters ( 88. 4％) . The training time for fixed pre-training parameters was 10 minutes,less than the strategy of fine-tuning pre-training parameters ( 16 hours ) and the strategy of randomly initializing parameters (24 hours). After the training data was reduced,the accuracy of the model obtained by the strategy of randomly initializing parameters decreased by 8. 6％ on average,while the accuracy of the transfer learning group decreased by 2. 5％ on average. Conclusions The proposed automated and novel DR detection algorithm based on fine-tune and NASNet structure maintains high accuracy in small sample dataset,is found to be robust,and effective for the preliminary diagnosis of DR.

Benign and malignant classification of renal occupancy lesions CT images using convolutional neural networks / 国际生物医学工程杂志

Objective To investigate the feasibility and application value of the benign and malignant classificational methods of renal occupying CT images based on convolutional neural networks (CNN). Methods An image omics method that can automatically learn the image features and classify CT images was used. Firstly, the CNN model obtained by large-scale natural image training was used to migrate the characteristics of the renal occupancy lesions CT images, and then the fine-tuning of the full connection layer was used to realize the benign and malignant classification of the images. Results The evaluation indexes of the VGG19 model were lower than ResNet50 and Inception V3, and the training result showed obvious overfitting. The accuracy, sensitivity and negative prediction values of the Inception V3 model was 93.8％, 99.5％ and 99.1％, respectively, which were higher than that of the ResNet50 model. Conclusions The benign and malignant classification of renal occupancy lesions CT images using CNN is a reasonable and feasible method, and the fine-tuned Inception V3 model has a better classification performance.

Tongue image classification method based on transfer learning and fully connected neural network / 第二军医大学学报

Objective To propose a classification method for small sample tongue images based on transfer learning and fully connected neural network, so as to solve the problems of large amount of data, high requirement of training equipment and long training time of deep learning in the classification of tongue images. Methods Effective features such as tongue points and lines of tongue images were extracted by the convolution Inception_v3 network after training on the massive data set of ImageNet. The above features were classified by the fully connected neural network, and the image knowledge acquired by the deep learning network was transferred to the tongue image recognition task, and then the tongue data set were used to train and test the efficiency of the network. Results Compared with the typical tongue image classification method such as K-nearest neighbor (KNN) algorithm, support vector machine (SVM) algorithm and convolutional neural network (CNN) deep learning method, the two methods (Inception_v3＋2NN and Inception_v3＋3NN) in our experiment had higher classification rates for tongue images, with the accuracy rates being 90.30% and 93.98%, respectively, and had shorter training time for the sample. Conclusion Compared with KNN algorithm, SVM algorithm and CNN deep learning method, the tongue image classification method based on transfer learning and fully connected neural network can effectively improve the accuracy rate of tongue image classification and shorten the training time.