Deep neural networks analysis of 18F-FDG PET imaging in postoperative patients with temporal lobe epilepsy / 中华核医学与分子影像杂志

Objective:To predict the short-term postoperative recurrence status of patients with refractory temporal lobe epilepsy (TLE) by analyzing preoperative 18F-FDG PET images and patients′ clinical characteristics based on deep residual neural network (ResNet). Methods:Retrospective analysis was conducted on preoperative 18F-FDG PET images and clinical data of 220 patients with refractory TLE (132 males and 88 females, age 23.0(20.0, 30.2) years)) in the First Affiliated Hospital of Jinan University between January 2014 and June 2020. ResNet was used to perform high-throughput feature extraction on preprocessed PET images and clinical features, and to perform a postoperative recurrence prediction task for differentiating patients with TLE. The predictive performance of ResNet model was evaluated by ROC curve analysis, and the AUC was compared with that of classical Cox proportional risk model using Delong test. Results:Based on PET images combined with clinical feature training, AUCs of the ResNet in predicting 12-, 24-, and 36-month postoperative recurrence were 0.895±0.073, 0.861±0.058 and 0.754±0.111, respectively, which were 0.717±0.093, 0.697±0.081 and 0.645±0.087 for Cox proportional hazards model respectively ( z values: -3.00, -2.98, -1.09, P values: 0.011, 0.018, 0.310). The ResNet showed best predictive effect for recurrence events within 12 months after surgery. Conclusion:The ResNet model is expected to be used in clinical practice for postoperative follow-up of patients with TLE, helping for risk stratification and individualized management of postoperative patients.

SPECT-MPI for Coronary Artery Disease: A deep learning approach

Background@#Worldwide, coronary artery disease (CAD) is a leading cause of mortality and morbidity and remains to be a top health priority in many countries. A non-invasive imaging modality for diagnosis of CAD such as single photon emission computed tomography-myocardial perfusion imaging (SPECT-MPI) is usually requested by cardiologists as it displays radiotracer distribution in the heart reflecting myocardial perfusion. The interpretation of SPECT-MPI is done visually by a nuclear medicine physician and is largely dependent on his clinical experience and showing significant inter-observer variability.@*Objective@#The aim of the study is to apply a deep learning approach in the classification of SPECT-MPI for perfusion abnormalities using convolutional neural networks (CNN).@*Methods@#A publicly available anonymized SPECT-MPI from a machine learning repository (https://www.kaggle.com/ selcankaplan/spect-mpi) was used in this study involving 192 patients who underwent stress-test-rest Tc99m MPI. An exploratory approach of CNN hyperparameter selection to search for optimum neural network model was utilized with particular focus on various dropouts (0.2, 0.5, 0.7), batch sizes (8, 16, 32, 64), and number of dense nodes (32, 64, 128, 256). The base CNN model was also compared with the commonly used pre-trained CNNs in medical images such as VGG16, InceptionV3, DenseNet121 and ResNet50. All simulations experiments were performed in Kaggle using TensorFlow 2.6.0., Keras 2.6.0, and Python language 3.7.10.@*Results@#The best performing base CNN model with parameters consisting of 0.7 dropout, batch size 8, and 32 dense nodes generated the highest normalized Matthews Correlation Coefficient at 0.909 and obtained 93.75% accuracy, 96.00% sensitivity, 96.00% precision, and 96.00% F1-score. It also obtained higher classification performance as compared to the pre-trained architectures. @*Conclusions@#The results suggest that deep learning approaches through the use of CNN models can be deployed by nuclear medicine physicians in their clinical practice to further augment their decision skills in the interpretation of SPECT-MPI tests. These CNN models can also be used as a dependable and valid second opinion that can aid physicians as a decision-support tool as well as serve as teaching or learning materials for the less-experienced physicians particularly those still in their training career. These highlights the clinical utility of deep learning approaches through CNN models in the practice of nuclear cardiology.

Quality control system based on artificial intelligence for improving imaging quality of chest CT / 中国医学影像技术

Objective To observe the value of quality control system based on artificial intelligence(AI)for improving imaging quality of chest CT.Methods Totally 1 726 CT images obtained from 415 patients were retrospectively collected,among which 1 414 images were used for convolutional neural network(CNN)training and the rest 312 images were used for validation.Precision,Recall,F1-Score,mean average precision(mAP)and intersection over union(IOU)of quality control system based on AI for chest CT scanning were calculated.Meanwhile,21 patients with unsatisfactory chest CT who would undergo re-examination were prospectively enrolled,and chest CT scanning with quality control system based on AI were performed.The results of 2 examinations were compared.Results Precision,Recall,F1-Score,mAP and IOU of quality control system based on AI for chest CT were all good.All 21 cases were diagnosed correctly with re-examination CT based on quality control system.Among 21 cases,the first CT misdiagnosed 19 cases,the displaying of the area,volume and display quality of pulmonary nodules were not significantly different,but the morphology,boundaries,spiny protrusions,vacuolar signs,inflatable bronchial signs of nodules as well as the thickened and twisted blood vessels were obviously different between 2 times examination.The first CT missed 1 case while correctly diagnosed 1 case.Conclusion The quality control system based on AI was helpful for improving imaging quality of chest CT and increasing diagnostic efficacy.

Ai based motion analysis software for sport and physical therapy assessment / Software de análisis de movimiento basado en ia para la evaluación del deporte y la fisioterapia / Software de análise de movimento baseado em ia para avaliação de esporte e fisioterapia

ABSTRACT Introduction: As the World Health Organization declared the novel coronavirus as a pandemic in March 2020, physical therapy is more difficult to execute, and social distancing is mandatory in the healthcare sector. Objective: In physical therapy, an online video analysis software that provides real-time graphic and numerical information about the patient's movement executions without direct personal contact would mean a significant improvement in eHealth treatment. Methods: We have developed a software layer on top of OpenPose human body position estimation software that can extract the time series of angles of arbitrary body parts using the output coordinates from OpenPose processing the data recorded by two cameras simultaneously. To validate the procedure of determining the joint angles using the Openpose software we have used the Kinovea software. Results: The comparison of the determined maximal knee angle in our and the Kinovea software, which is widely used in biomechanical measurements, was not significantly different (2.03±1.06°, p<0.05) Conclusion: This indicates, that the developed software can calculate the appropriate joint angles with the accuracy that physiotherapy treatments require. As, to our knowledge no such software yet exists, with the help of this software development, therapists could control and correct the exercises in real-time, and also from a distance, and physical therapy effectiveness could be increased. Level of Evidence II; Experimental, comparative.

RESUMEN Introducción: Como la Organización Mundial de la Salud declaró el nuevo coronavirus como una pandemia en marzo de 2020, la fisioterapia es más difícil de ejecutar, el distanciamiento social es obligatorio en el sector de la salud. Objetivo: En la práctica de fisioterapia un software de análisis de vídeo online que proporcione información gráfica y numérica en tiempo real sobre las ejecuciones de movimiento del paciente sin contacto personal directo supondría una mejora significativa en el tratamiento de la eSalud. Métodos: Fue desarrollado una capa de software sobre el software de estimación de posición del cuerpo humano OpenPose que puede extraer la serie temporal de ángulos de partes arbitrarias del cuerpo utilizando las coordenadas de salida de OpenPose procesando los datos registrados por dos cámaras simultáneamente. Para validar el procedimiento de determinación de los ángulos articulares mediante el software Openpose fue utilizado el software Kinovea. Resultados: La comparación del ángulo máximo de rodilla determinado en nuestro software y Kinovea, que es ampliamente utilizado en mediciones biomecánicas, no fue significativamente diferente (2,03±1,06°, p<0,05). Conclusión: Esto indica que el software desarrollado puede calcular los ángulos articulares adecuados con la precisión que requieren los tratamientos de fisioterapia. Dado que aún no existe dicho software, con la ayuda de este desarrollo de software, los terapeutas podrían controlar y corregir los ejercicios en tiempo real, y también a distancia, y se podría aumentar la eficacia de la fisioterapia. Nivel de Evidencia II; Experimental, comparativo.

RESUMO Introdução: Como a Organização Mundial da Saúde declarou o novo coronavírus como pandemia em março de 2020, a fisioterapia é mais difícil de executar, o distanciamento social é obrigatório no setor de saúde. Objetivo: Na prática da fisioterapia, um software de análise de vídeo online que fornece informações gráficas e numéricas em tempo real sobre as execuções de movimento do paciente sem contato pessoal direto significaria uma melhora significativa no tratamento eHealth. Métodos: Desenvolveu-se uma camada de software em cima do software de estimativa de posição do corpo humano OpenPose que pode extrair as séries temporais de ângulos de partes do corpo arbitrárias usando as coordenadas de saída do OpenPose processando os dados gravados por duas câmeras simultaneamente. Para validar o procedimento de determinação dos ângulos articulares utilizando o software Openpose utilizou-se o software Kinovea. Resultados: A comparação do ângulo máximo do joelho determinado em nosso e no software Kinovea, amplamente utilizado em medidas biomecânicas, não foi significativamente diferente (2,03±1,06°, p<0,05) Conclusão: Isso indica que o software desenvolvido pode calcular os ângulos articulares adequados com a precisão que os tratamentos de fisioterapia exigem. Como esse software ainda não existe, com a ajuda do desenvolvimento desse software, os terapeutas puderam controlar e corrigir os exercícios em tempo real, e também à distância, aumentando a eficácia da fisioterapia. Nível de Evidência II; Experimental, comparativo.

Artificial intelligence to predict bed bath time in Intensive Care Units / Inteligencia artificial para predecir la hora del baño en cama en Unidades de Cuidados Intensivos / Inteligência artificial para predição do tempo de banho no leito em Unidades de Terapia Intensiva

ABSTRACT Objectives: to assess the predictive performance of different artificial intelligence algorithms to estimate bed bath execution time in critically ill patients. Methods: a methodological study, which used artificial intelligence algorithms to predict bed bath time in critically ill patients. The results of multiple regression models, multilayer perceptron neural networks and radial basis function, decision tree and random forest were analyzed. Results: among the models assessed, the neural network model with a radial basis function, containing 13 neurons in the hidden layer, presented the best predictive performance to estimate the bed bath execution time. In data validation, the squared correlation between the predicted values and the original values was 62.3%. Conclusions: the neural network model with radial basis function showed better predictive performance to estimate bed bath execution time in critically ill patients.

RESUMEN Objetivos: evaluar el rendimiento predictivo de diferentes algoritmos de inteligencia artificial para estimar el tiempo de ejecución del baño en cama en pacientes críticos. Métodos: estudio metodológico, que utilizó algoritmos de inteligencia artificial para predecir el tiempo de baño en cama en pacientes críticos. Se analizaron los resultados de modelos de regresión múltiple, redes neuronales perceptrón multicapa y función de base radial, árbol de decisión y random forest. Resultados: entre los modelos evaluados, el modelo de red neuronal con función de base radial, que contiene 13 neuronas en la capa oculta, presentó el mejor desempeño predictivo para estimar el tiempo de ejecución del baño en cama. En la validación de datos, la correlación al cuadrado entre los valores predichos y los valores originales fue del 62,3%. Conclusiones: el modelo de red neuronal con función de base radial mostró mejor rendimiento predictivo para estimar el tiempo de ejecución del baño en cama en pacientes críticos.

RESUMO Objetivos: avaliar a performance preditiva de diferentes algoritmos de inteligência artificial para estimar o tempo de execução do banho no leito em pacientes críticos. Métodos: estudo metodológico, que utilizou algoritmos de inteligência artificial para predizer o tempo de banho no leito em pacientes críticos. Foram analisados os resultados dos modelos de regressão múltipla, redes neurais perceptron multicamadas e função de base radial, árvore de decisão e random forest. Resultados: entre os modelos avaliados, o modelo de rede neural com função de base radial, contendo 13 neurônios na camada oculta, apresentou melhor performance preditiva para estimar o tempo de execução do banho no leito. Na validação dos dados, o quadrado da correlação entre os valores preditos e os valores originais foi de 62,3%. Conclusões: o modelo de rede neural com função de base radial apresentou melhor performance preditiva para estimar o tempo de execução do banho no leito em pacientes críticos.

Performance analysis of pretrained convolutional neural network models for ophthalmological disease classification / Análise de desempenho de modelos de rede neural convolucional pré-treinados para classificação de doenças oftalmológicas

ABSTRACT Purpose: This study aimed to evaluate the classification performance of pretrained convolutional neural network models or architectures using fundus image dataset containing eight disease labels. Methods: A publicly available ocular disease intelligent recognition database has been used for the diagnosis of eight diseases. This ocular disease intelligent recognition database has a total of 10,000 fundus images from both eyes of 5,000 patients for the following eight diseases: healthy, diabetic retinopathy, glaucoma, cataract, age-related macular degeneration, hypertension, myopia, and others. Ocular disease classification performances were investigated by constructing three pretrained convolutional neural network architectures including VGG16, Inceptionv3, and ResNet50 models with adaptive moment optimizer. These models were implemented in Google Colab, which made the task straight-forward without spending hours installing the environment and supporting libraries. To evaluate the effectiveness of the models, the dataset was divided into 70%, 10%, and 20% for training, validation, and testing, respectively. For each classification, the training images were augmented to 10,000 fundus images. Results: ResNet50 achieved an accuracy of 97.1%; sensitivity, 78.5%; specificity, 98.5%; and precision, 79.7%, and had the best area under the curve and final score to classify cataract (area under the curve = 0.964, final score = 0.903). By contrast, VGG16 achieved an accuracy of 96.2%; sensitivity, 56.9%; specificity, 99.2%; precision, 84.1%; area under the curve, 0.949; and final score, 0.857. Conclusions: These results demonstrate the ability of the pretrained convolutional neural network architectures to identify ophthalmological diseases from fundus images. ResNet50 can be a good architecture to solve problems in disease detection and classification of glaucoma, cataract, hypertension, and myopia; Inceptionv3 for age-related macular degeneration, and other disease; and VGG16 for normal and diabetic retinopathy.

RESUMO Objetivo: Avaliar o desempenho de classificação de modelos ou arquiteturas de rede neural convolucional pré--treinadas usando um conjunto de dados de imagem de fundo de olho contendo oito rótulos de doenças diferentes. Métodos: Neste artigo, o conjunto de dados de reconhecimento inteligente de doenças oculares publicamente disponível foi usado para o diagnóstico de oito rótulos de doenças diferentes. O banco de dados de reconhecimento inteligente de doenças oculares tem um total de 10.000 imagens de fundo de olho de ambos os olhos de 5.000 pacientes para oito categorias que contêm rótulos saudáveis, retinopatia diabética, glaucoma, catarata, degeneração macular relacionada à idade, hipertensão, miopia, outros. Investigamos o desempenho da classificação de doenças oculares construindo três arquiteturas de rede neural convolucional pré-treinadas diferentes, incluindo os modelos VGG16, Inceptionv3 e ResNet50 com otimizador de Momento Adaptativo. Esses modelos foram implementados no Google Colab o que facilitou a tarefa sem gastar horas instalando o ambiente e suportando bibliotecas. Para avaliar a eficácia dos modelos, o conjunto de dados é dividido em 70% para treinamento, 10% para validação e os 20% restantes utilizados para teste. As imagens de treinamento foram expandidas para 10.000 imagens de fundo de olho para cada tal. Resultados: Observou-se que o modelo ResNet50 alcançou acurácia de 97,1%, sensibilidade de 78,5%, especificidade de 98,5% e precisão de 79,7% e teve a melhor área sob a curva e pontuação final para classificar a categoria da catarata (área sob a curva=0,964, final=0,903). Em contraste, o modelo VGG16 alcançou uma precisão de 96,2%, sensibilidade de 56,9%, especificidade de 99,2% e precisão de 84,1%, área sob a curva 0,949 e pontuação final de 0,857. Conclusão: Esses resultados demonstram a capacidade das arquiteturas de rede neural convolucional pré-treinadas em identificar doenças oftalmológicas a partir de imagens de fundo de olho. ResNet50 pode ser uma boa solução para resolver problemas na detecção e classificação de doenças como glaucoma, catarata, hipertensão e miopia; Inceptionv3 para degeneração macular relacionada à idade e outras doenças; e VGG16 para retinopatia normal e diabética.

Machine learning in predicting severe acute respiratory infection outbreaks / Aprendizado de máquina na previsão de surtos de síndrome respiratória aguda grave / Aprendizaje automático en la predicción de brotes de síndrome respiratorio agudo grave

Abstract: Severe acute respiratory infection (SARI) outbreaks occur annually, with seasonal peaks varying among geographic regions. Case notification is important to prepare healthcare networks for patient attendance and hospitalization. Thus, health managers need adequate resource planning tools for SARI seasons. This study aims to predict SARI outbreaks based on models generated with machine learning using SARI hospitalization notification data. In this study, data from the reporting of SARI hospitalization cases in Brazil from 2013 to 2020 were used, excluding SARI cases caused by COVID-19. These data were prepared to feed a neural network configured to generate predictive models for time series. The neural network was implemented with a pipeline tool. Models were generated for the five Brazilian regions and validated for different years of SARI outbreaks. By using neural networks, it was possible to generate predictive models for SARI peaks, volume of cases per season, and for the beginning of the pre-epidemic period, with good weekly incidence correlation (R2 = 0.97; 95%CI: 0.95-0.98, for the 2019 season in the Southeastern Brazil). The predictive models achieved a good prediction of the volume of reported cases of SARI; accordingly, 9,936 cases were observed in 2019 in Southern Brazil, and the prediction made by the models showed a median of 9,405 (95%CI: 9,105-9,738). The identification of the period of occurrence of a SARI outbreak is possible using predictive models generated with neural networks and algorithms that employ time series.

Resumo: Surtos de síndrome respiratória aguda grave (SRAG) ocorrem anualmente, com picos sazonais variando entre regiões geográficas. A notificação dos casos é importante para preparar as redes de atenção à saúde para o atendimento e internação dos pacientes. Portanto, os gestores de saúde precisam ter ferramentas adequadas de planejamento de recursos para as temporadas de SRAG. Este estudo tem como objetivo prever surtos de SRAG com base em modelos gerados com aprendizado de máquina usando dados de internação por SRAG. Foram incluídos dados sobre casos de hospitalização por SRAG no Brasil de 2013 a 2020, excluindo os casos causados pela COVID-19. Estes dados foram preparados para alimentar uma rede neural configurada para gerar modelos preditivos para séries temporais. A rede neural foi implementada com uma ferramenta de pipeline. Os modelos foram gerados para as cinco regiões brasileiras e validados para diferentes anos de surtos de SRAG. Com o uso de redes neurais, foi possível gerar modelos preditivos para picos de SRAG, volume de casos por temporada e para o início do período pré-epidêmico, com boa correlação de incidência semanal (R2 = 0,97; IC95%: 0,95-0,98, para a temporada de 2019 na Região Sudeste). Os modelos preditivos obtiveram uma boa previsão do volume de casos notificados de SRAG; dessa forma, foram observados 9.936 casos em 2019 na Região Sul, e a previsão feita pelos modelos mostrou uma mediana de 9.405 (IC95%: 9.105-9.738). A identificação do período de ocorrência de um surto de SRAG é possível por meio de modelos preditivos gerados com o uso de redes neurais e algoritmos que aplicam séries temporais.

Resumen: Brotes de síndrome respiratorio agudo grave (SRAG) ocurren todos los años, con picos estacionales que varían entre regiones geográficas. La notificación de los casos es importante para preparar las redes de atención a la salud para el cuidado y hospitalización de los pacientes. Por lo tanto, los gestores de salud deben tener herramientas adecuadas de planificación de recursos para las temporadas de SRAG. Este estudio tiene el objetivo de predecir brotes de SRAG con base en modelos generados con aprendizaje automático utilizando datos de hospitalización por SRAG. Se incluyeron datos sobre casos de hospitalización por SRAG en Brasil desde 2013 hasta 2020, salvo los casos causados por la COVID-19. Se prepararon estos datos para alimentar una red neural configurada para generar modelos predictivos para series temporales. Se implementó la red neural con una herramienta de canalización. Se generaron los modelos para las cinco regiones brasileñas y se validaron para diferentes años de brotes de SRAG. Con el uso de redes neurales, se pudo generar modelos predictivos para los picos de SRAG, el volumen de casos por temporada y para el inicio del periodo pre-epidémico, con una buena correlación de incidencia semanal (R2 = 0,97; IC95%: 0,95-0,98, para la temporada de 2019 en la Región Sudeste). Los modelos predictivos tuvieron una buena predicción del volumen de casos notificados de SRAG; así, se observaron 9.936 casos en 2019 en la Región Sur, y la predicción de los modelos mostró una mediana de 9.405 (IC95%: 9.105-9.738). La identificación del periodo de ocurrencia de un brote de SRAG es posible a través de modelos predictivos generados con el uso de redes neurales y algoritmos que aplican series temporales.

Além das ferramentas tradicionais: explorando redes neurais convolucionais como modelos prognósticos inovadores no adenocarcinoma ductal pancreático / Beyond traditional tools: exploring convolutional neural networks as innovative prognostic models in pancreatic ductal adenocarcinoma

ABSTRACT Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive and lethal form of cancer with limited prognostic accuracy using traditional factors. This has led to the exploration of innovative prognostic models, including convolutional neural networks (CNNs), in PDAC. CNNs, a type of artificial intelligence algorithm, have shown promise in various medical applications, including image analysis and pattern recognition. Their ability to extract complex features from medical images makes them suitable for improving prognostication in PDAC. However, implementing CNNs in clinical practice poses challenges, such as data availability and interpretability. Future research should focus on multi-center studies, integrating multiple data modalities, and combining CNN outputs with biomarker panels. Collaborative efforts and patient autonomy should be considered to ensure the ethical implementation of CNN-based prognostic models. Further validation and optimisation of CNN-based models are necessary to enhance their reliability and clinical utility in PDAC prognostication.

RESUMO Contexto O adenocarcinoma ductal pancreático (ACDP) é uma forma de câncer altamente agressiva e letal com precisão prognóstica limitada usando fatores tradicionais. Isso levou à exploração de modelos prognósticos inovadores, incluindo redes neurais convolucionais (CNNs), no ACDP. As CNNs, um tipo de algoritmo de inteligência artificial, mostraram promessa em várias aplicações médicas, incluindo análise de imagem e reconhecimento de padrões. Sua capacidade de extrair características complexas de imagens médicas as torna adequadas para melhorar o prognóstico no ACDP. No entanto, a implementação de CNNs na prática clínica apresenta desafios, como a disponibilidade de dados e a interpretabilidade. Pesquisas futuras devem se concentrar em estudos multicêntricos, integrando múltiplas modalidades de dados e combinando saídas de CNN com painéis de biomarcadores. Esforços colaborativos e autonomia do paciente devem ser considerados para garantir a implementação ética de modelos prognósticos baseados em CNN. Mais validação e otimização de modelos baseados em CNN são necessárias para aumentar sua confiabilidade e utilidade clínica na prognostico do ACDP.

Natural language processing in the classification of radiology reports in benign gallbladder diseases / O processamento de língua natural permite a classificação correta de laudos radiológicos em doenças benignas da vesícula biliar

Abstract Objective: To develop a natural language processing application capable of automatically identifying benign gallbladder diseases that require surgery, from radiology reports. Materials and Methods: We developed a text classifier to classify reports as describing benign diseases of the gallbladder that do or do not require surgery. We randomly selected 1,200 reports describing the gallbladder from our database, including different modalities. Four radiologists classified the reports as describing benign disease that should or should not be treated surgically. Two deep learning architectures were trained for classification: a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) network. In order to represent words in vector form, the models included a Word2Vec representation, with dimensions of 300 or 1,000. The models were trained and evaluated by dividing the dataset into training, validation, and subsets (80/10/10). Results: The CNN and BiLSTM performed well in both dimensional spaces. For the 300- and 1,000-dimensional spaces, respectively, the F1-scores were 0.95945 and 0.95302 for the CNN model, compared with 0.96732 and 0.96732 for the BiLSTM model. Conclusion: Our models achieved high performance, regardless of the architecture and dimensional space employed.

Resumo Objetivo: Desenvolver uma aplicação de processamento de linguagem natural capaz de identificar automaticamente doenças cirúrgicas benignas da vesícula biliar a partir de laudos radiológicos. Materiais e Métodos: Desenvolvemos um classificador de texto para classificar laudos como contendo ou não doenças cirúrgicas benignas da vesícula biliar. Selecionamos aleatoriamente 1.200 laudos com descrição da vesícula biliar de nosso banco de dados, incluindo diferentes modalidades. Quatro radiologistas classificaram os laudos como doença benigna cirúrgica ou não. Duas arquiteturas de aprendizagem profunda foram treinadas para a classificação: a rede neural convolucional (convolutional neural network - CNN) e a memória longa de curto prazo bidirecional (bidirectional long short-term memory - BiLSTM). Para representar palavras de forma vetorial, os modelos incluíram uma representação Word2Vec, com dimensões variando de 300 a 1000. Os modelos foram treinados e avaliados por meio da divisão do conjunto de dados entre treinamento, validação e teste (80/10/10). Resultados: CNN e BiLSTM tiveram bom desempenho em ambos os espaços dimensionais. Relatamos para 300 e 1000 dimensões, respectivamente, as pontuações F1 de 0,95945 e 0,95302 para o modelo CNN e de 0,96732 e 0,96732 para a BiLSTM. Conclusão: Nossos modelos alcançaram alto desempenho, independentemente de diferentes arquiteturas e espaços dimensionais.

Tomografía fotoacústica y Deep Learning en aplicacionesmédicas

En las últimas décadas, las imágenes fotoacústicas han demostrado su eficacia en el apoyo al diagnóstico de algunas enfermedades, así como en la investigación médica, ya que a través de ellas es posible obtener información del cuerpo humano con características específicas y profundidad de penetración, desde 1 cm hasta 6 cm dependiendo en gran medida del tejido estudiado, además de una buena resolución. Las imágenes fotoacústicas son comparativamente jóvenes y emergentes y prometen mediciones en tiempo real, con procedimientos no invasivos y libres de radiación. Por otro lado, aplicar Deep Learning a imágenes fotoacústicas permite gestionar datos y transformarlos en información útil que genere conocimiento. Estas aplicaciones poseen ventajas únicas que facilitan la aplicación clínica. Se considera que con estas técnicas se pueden proporcionar diagnósticos médicos confiables. Es por eso que el objetivo de este artículo es proporcionar un panorama general de los casos donde se combina el Deep Learning con técnicas fotoacústicas.

In recent decades, photoacoustic imaging has proven its effectiveness in supporting the diagnosis of some diseases as well as in medical research, since through them it is possible to obtain information of the human body with specific characteristics and depth of penetration, from 1 cm to 6 cm depending largely on the tissue studied, in addition to a good resolution. Photoacoustic imaging is comparatively young and emerging and promises real-time measurements, with non-invasive and radiation-free procedures. On the other hand, applying Deep Learning to photoacoustic images allows managing data and transforming them into useful information that generates knowledge. These applications have unique advantages that facilitate clinical application. It may be possible with these techniques to provide reliable medical diagnoses. That is why the aim of this article is to provide an overview of cases combining Deep Learning with photoacoustic techniques.

Development of a convolutional neural network for diagnosing osteoarthritis, trained with knee radiographs from the ELSA-Brasil Musculoskeletal / Desenvolvimento de rede neural convolucional para o diagnóstico radiográfico de osteoartrite dos joelhos no ELSA-Brasil Musculoesquelético

Abstract Objective: To develop a convolutional neural network (CNN) model, trained with the Brazilian "Estudo Longitudinal de Saúde do Adulto Musculoesquelético" (ELSA-Brasil MSK, Longitudinal Study of Adult Health, Musculoskeletal) baseline radiographic examinations, for the automated classification of knee osteoarthritis. Materials and Methods: This was a cross-sectional study carried out with 5,660 baseline posteroanterior knee radiographs from the ELSA-Brasil MSK database (5,660 baseline posteroanterior knee radiographs). The examinations were interpreted by a radiologist with specific training, and the calibration was as established previously. Results: The CNN presented an area under the receiver operating characteristic curve of 0.866 (95% CI: 0.842-0.882). The model can be optimized to achieve, not simultaneously, maximum values of 0.907 for accuracy, 0.938 for sensitivity, and 0.994 for specificity. Conclusion: The proposed CNN can be used as a screening tool, reducing the total number of examinations evaluated by the radiologists of the study, and as a double-reading tool, contributing to the reduction of possible interpretation errors.

Resumo Objetivo: Desenvolver um modelo computacional - rede neural convolucional (RNC) - treinado com radiografias da linha de base do Estudo Longitudinal de Saúde do Adulto Musculoesquelético (ELSA-Brasil Musculoesquelético), para a classificação automática de osteoartrite dos joelhos. Materiais e Métodos: Trata-se de um estudo transversal abrangendo todos os exames da linha de base do ELSA-Brasil Musculoesquelético (5.660 radiografias dos joelhos em incidência posteroanterior). Os exames foram interpretados por médico radiologista com treinamento específico e calibração previamente publicada. Resultados: A RNC desenvolvida apresentou área sob a curva característica de operação do receptor de 0,866 (IC 95%: 0,842-0,882). O modelo pode ser calibrado para alcançar, não simultaneamente, valores máximos de 0,907 para acurácia, 0,938 para sensibilidade e 0,994 para especificidade. Conclusão: A RNC desenvolvida pode ser utilizada como ferramenta de triagem, reduzindo o número total de exames avaliados pelos radiologistas do estudo, e/ou como ferramenta de segunda leitura, contribuindo com a redução de possíveis erros de interpretação.

Artificial intelligence: Its role in dermatopathology

Artificial intelligence (AI), a major frontier in the field of medical research, can potentially lead to a paradigm shift in clinical practice. A type of artificial intelligence system known as convolutional neural network points to the possible utility of deep learning in dermatopathology. Though pathology has been traditionally restricted to microscopes and glass slides, recent advancement in digital pathological imaging has led to a transition making it a potential branch for the implementation of artificial intelligence. The current application of artificial intelligence in dermatopathology is to complement the diagnosis and requires a well-trained dermatopathologist’s guidance for better designing and development of deep learning algorithms. Here we review the recent advances of artificial intelligence in dermatopathology, its applications in disease diagnosis and in research, along with its limitations and future potential

Predicción de la evapotranspiración de referencia utilizando redes neuronales artificiales / Reference evapotranspiration prediction using Artificial Neural Networks

RESUMEN La evapotranspiración de referencia (ETo) es una variable hidrológica de gran importancia en el manejo del riego. Su estimación se realiza con la ecuación de Penman-Montieth (PM), que requiere de muchas variables meteorológicas, las cuales, a veces, no se encuentran disponibles. Dado que la ETo es una variable no lineal y compleja, en los últimos años han surgido métodos alternativos para su estimación, como las redes neuronales artificiales (RNA). El objetivo del presente trabajo fue estimar la evapotranspiración de referencia (ETo) usando la ecuación de Penman-Montieth, a fin de desarrollar modelos de redes neuronales artificiales (RNA) que permitan predecir la ETo en regiones con información climatológica limitada, y su vez comparar el desempeño de tres modelos de RNA: FFNN, ERNN y NARX. Se utilizó información diaria durante el periodo 1 de enero de 2007 al 31 de diciembre de 2018, de las estaciones meteorológicas ENP8 y ENP4 de la CDMX. Se realizó un análisis de correlación y el análisis de sensibilidad de Garson para estudiar 2 casos (red estática FFNN y redes dinámicas: ERNN y NARX) usando 3 modelos de RNA: 1) RNA con 6 entradas: radiación solar (Rad), temperatura máxima y mínima (Tmax, Tmin), humedad relativa máxima y mínima (HRmax, HRmin) y velocidad del viento (u); y 2) RNA con 2 entradas (Rad y Tmax). La variable de salida fue la ETo calculada con la ecuación de PM. En todos los casos, las 3 RNA fueron muy parecidas, la diferencia más notable es que las redes dinámicas (ERNN y NARX) requieren de menor número de iteraciones para llegar al desempeño óptimo. Las RNA entrenadas, únicamente con Rad y Tmax como entradas, fueron capaces de predecir la ETo en el largo plazo, durante 440 d, en otra estación meteorológica cercana (ENP4), con eficiencias mayores al 90 %.

ABSTRACT Reference evapotranspiration (ETo) is a hydrological variable of great importance in irrigation management. Its estimation is carried out with the Penman-Montieth (PM) equation that requires many meteorological variables and that are sometimes not available. Since ETo is a nonlinear and complex variable, in recent years alternative methods have emerged for its estimation, such as artificial neural networks (ANN). The objective of this work was to estimate the reference evapotranspiration (ETo) using the Penman-Montieth equation, in order to develop artificial neural network (ANN) models that allow ETo to be predicted in regions with limited climatological information, and in turn to compare the performance of three RNA models: FFNN, ERNN and NARX. Daily informtion was used during the January 1, 2007 to December 31, 2018 period, for the ENP8 and ENP4 meteorological stations in Mexico city. Based on the correlation analysis and the Garson sensitivity analysis, 2 cases were studied for the 3 ANN models: 1) ANN with 6 inputs: solar radiation (Rad), maximum and minimum temperature (Tmax, Tmin), maximum and minimum relative humidity (RHmax, RHmin), and wind speed (u), and 2) RNA with 2 inputs (Rad and Tmax). The output variable was the ETo, calculated with the PM equation. In all cases, the performance of the 3 ANNs was very similar. The most notable difference is that the dynamic networks (ERNN and NARX) require fewer iterations to achieve the optimum performance. ANNs trained only with radiation and maximum temperature as inputs were able to predict a long-term ETo for 440 at another nearby meteorological station (ENP4), with efficiencies greater than 90 %.

Red neural artificial para predecir factores de riesgo asociados a complicaciones postoperatorias secundarias al tratamiento del neumotórax / Artificial neural network to predict risk factors associated with postoperative complications secondary to pneumothorax treatment

Introducción. Debido a la ausencia de modelos predictivos estadísticamente significativos enfocados a las complicaciones postoperatorias en el manejo quirúrgico del neumotórax, desarrollamos un modelo, utilizando redes neurales, que identifica las variables independientes y su importancia para reducir la incidencia de complicaciones. Métodos. Se realizó un estudio retrospectivo en un centro asistencial, donde se incluyeron 106 pacientes que requirieron manejo quirúrgico de neumotórax. Todos fueron operados por el mismo cirujano. Se desarrolló una red neural artificial para manejo de datos con muestras limitadas; se optimizaron los datos y cada algoritmo fue evaluado de forma independiente y mediante validación cruzada, para obtener el menor error posible y la mayor precisión con el menor tiempo de respuesta. Resultados. Las variables de mayor importancia según su peso en el sistema de decisión de la red neural (área bajo la curva 0,991) fueron el abordaje por toracoscopia video asistida (OR 1,131), el uso de pleurodesis con talco (OR 0,994) y el uso de autosuturas (OR 0,792; p<0,05). Discusión. En nuestro estudio, los principales predictores independientes asociados a mayor riesgo de complicaciones fueron el neumotórax de etiología secundaria y el neumotórax recurrente. Adicionalmente, confirmamos que las variables asociadas a reducción de riesgo de complicaciones postoperatorias tuvieron significancia estadística. Conclusión. Identificamos la toracoscopia video asistida, el uso de autosuturas y la pleurodesis con talco como posibles variables asociadas a menor riesgo de complicaciones. Se plantea la posibilidad de desarrollar una herramienta que facilite y apoye la toma de decisiones, por lo cual es necesaria la validación externa en estudios prospectivos

Introduction. Due to the absence of statistically significant predictive models focused on postoperative complications in the surgical management of pneumothorax, we developed a model using neural networks that identify the independent variables and their importance in reducing the incidence of postoperative complications. Methods. A retrospective single-center study was carried out, where 106 patients who required surgical management of pneumothorax were included. All patients were operated by the same surgeon. An artificial neural network was developed to manage data with limited samples. The data is optimized and each algorithm is evaluated independently and through cross-validation to obtain the lowest possible error and the highest precision with the shortest response time. Results. The most important variables according to their weight in the decision system of the neural network (AUC 0.991) were the approach via video-assisted thoracoscopy (OR 1.131), use of pleurodesis with powder talcum (OR 0.994) and use of autosutures (OR 0.792, p<0.05). Discussion. In our study, the main independent predictors associated with a higher risk of complications are pneumothorax of secondary etiology and recurrent pneumothorax. Additionally, we confirm that the variables associated with a reduction in the risk of postoperative complications have statistical significance. Conclusion. We identify video-assisted thoracoscopy, use of autosuture and powder talcum pleurodesis as possible variables associated with a lower risk of complications and raise the possibility of developing a tool that facilitates and supports decision-making, for which external validation in prospective studies is necessary

Predictive model for urosepsis in patients with Upper Urinary Tract Calculi based on ultrasonography and urinalysis using artificial intelligence learning

ABSTRACT Purpose To construct a predicting model for urosepsis risk for patients with upper urinary tract calculi based on ultrasound and urinalysis. Materials and Methods A retrospective study was conducted in patients with upper urinary tract calculi admitted between January 2016 and January 2020. The patients were randomly grouped into the training and validation sets. The training set was used to identify the urosepsis risk factors and construct a risk prediction model based on ultrasound and urinalysis. The validation set was used to test the performance of the artificial neural network (ANN). Results Ultimately, 1716 patients (10.8% cases and 89.2% control) were included. Eight variables were selected for the model: sex, age, body temperature, diabetes history, urine leukocytes, urine nitrite, urine glucose, and degree of hydronephrosis. The area under the receiver operating curve in the validation and training sets was 0.945 (95% CI: 0.903-0.988) and 0.992 (95% CI: 0.988-0.997), respectively. Sensitivity, specificity, and Yuden index of the validation set (training set) were 80.4% (85.9%), 98.2% (99.0%), and 0.786 (0.849), respectively. Conclusions A preliminary screening model for urosepsis based on ultrasound and urinalysis was constructed using ANN. The model could provide risk assessments for urosepsis in patients with upper urinary tract calculi.

Identification of Chinese Herb Pieces Based on YOLOv4 / 中国实验方剂学杂志

Chinese herbal piece is an important component of the traditional Chinese medicine （TCM） system， and identifying their quality and grading can promote the development and utilization of Chinese herbal pieces. Utilizing deep learning for intelligent identification of Chinese herbal pieces can save time， effort， and cost， while also reasonably avoiding the constraints of human subjectivity， providing a guarantee for efficient identification of Chinese herbal pieces. In this study， a dataset containing 108 kinds of Chinese herbal pieces （14 058 images） was constructed，the basic YOLOv4 algorithm was employed to identify the 108 kinds of Chinese herbal pieces of our database The mean average precision （mAP） of the developed basic YOLOv4 model reached 85.3%. In addition， the receptive field block was introduced into the neck network of YOLOv4 algorithm， and the improved YOLOv4 algorithm was used to identify Chinese herbal pieces. The mAPof the improved YOLOv4 model achieved 88.7%， the average precision of 80 kinds of decoction pieces exceeded 80%， the average precision of 48 kinds of decoction pieces exceeded 90%. These results indicate that adding the receptive field module can help to some extent in the identification of Chinese herbal medicine pieces with different sizes and small volumes. Finally， the average precision of each kind of Chinese herbal medicine piece by the improved YOLOv4 model was further analyzed. Through in-depth analysis of the original images of Chinese herbal medicine pieces with low prediction average precision， it was clarified that the quantity and quality of original images of Chinese herbal medicine pieces are key to performing intelligent object detection. The improved YOLOv4 model constructed in this study can be used for the rapid identification of Chinese herbal pieces， and also provide reference guidance for the manual authentication of Chinese herbal medicine decoction pieces.

Application of Novel Down-sampling Method in Retinal Vessel Segmentation / 中国医疗器械杂志

Accurate segmentation of retinal blood vessels is of great significance for diagnosing, preventing and detecting eye diseases. In recent years, the U-Net network and its various variants have reached advanced level in the field of medical image segmentation. Most of these networks choose to use simple max pooling to down-sample the intermediate feature layer of the image, which is easy to lose part of the information, so this study proposes a simple and effective new down-sampling method Pixel Fusion-pooling (PF-pooling), which can well fuse the adjacent pixel information of the image. The down-sampling method proposed in this study is a lightweight general module that can be effectively integrated into various network architectures based on convolutional operations. The experimental results on the DRIVE and STARE datasets show that the F1-score index of the U-Net model using PF-pooling on the STARE dataset improved by 1.98%. The accuracy rate is increased by 0.2%, and the sensitivity is increased by 3.88%. And the generalization of the proposed module is verified by replacing different algorithm models. The results show that PF-pooling has achieved performance improvement in both Dense-UNet and Res-UNet models, and has good universality.

An Intelligent Method for Optimizing Large-volume Medical Image Transmission / 中国医疗器械杂志

In order to alleviate the conflict between medical supply and demand, and to improve the efficiency of medical image transmission, this study proposes an intelligent method for large-volume medical image transmission. This method extracts and generates keyword pairs by analyzing medical diagnostic reports, and uses a 3D-UNet to segment original image data into various sub-area based on its anatomy structure. Then, the sub-areas are scored through keyword pairs and preset scoring criteria, and transmitted to user frontend in the order of prioritization score. Experiments show that this method can fulfill physicians' requirements of radiology reading and diagnosis with only ten percent of data transmitted, which efficiently optimized traditional transmission procedures.

Kinome-wide polypharmacology profiling of small molecules by multi-task graph isomorphism network approach

Prediction of the interactions between small molecules and their targets play important roles in various applications of drug development, such as lead discovery, drug repurposing and elucidation of potential drug side effects. Therefore, a variety of machine learning-based models have been developed to predict these interactions. In this study, a model called auxiliary multi-task graph isomorphism network with uncertainty weighting (AMGU) was developed to predict the inhibitory activities of small molecules against 204 different kinases based on the multi-task Graph Isomorphism Network (MT-GIN) with the auxiliary learning and uncertainty weighting strategy. The calculation results illustrate that the AMGU model outperformed the descriptor-based models and state-of-the-art graph neural networks (GNN) models on the internal test set. Furthermore, it also exhibited much better performance on two external test sets, suggesting that the AMGU model has enhanced generalizability due to its great transfer learning capacity. Then, a naïve model-agnostic interpretable method for GNN called edges masking was devised to explain the underlying predictive mechanisms, and the consistency of the interpretability results for 5 typical epidermal growth factor receptor (EGFR) inhibitors with their structure‒activity relationships could be observed. Finally, a free online web server called KIP was developed to predict the kinome-wide polypharmacology effects of small molecules (http://cadd.zju.edu.cn/kip).